An Exploratory, Forward-Only Paper Trading Study by Nathan Smith一项由 Nathan Smith 进行的探索性、单向模拟交易研究
Large language models (LLMs) are increasingly being applied to financial tasks; however, systematic research on LLMs acting as autonomous decision-making agents remains limited. This evaluation presents a paper-trading study of ChatGPT acting as a portfolio decision-maker within the U.S. listed micro-cap equity market.大语言模型（LLM）正被引入金融任务。但作为自主决策代理，系统的研究还很少。本次评估针对 ChatGPT 在美股微盘股市场的投资决策，进行了一项模拟交易研究。
Over the six-month experimental period, the model was prompted using daily trading updates and weekly portfolio evaluations. ChatGPT had complete control over portfolio decisions, while human input was strictly limited to prompting and trade execution.实验周期六个月。每天给模型喂交易更新，每周做组合评估。模型全权决定买卖，人只负责输入提示和执行交易。
The evaluation found that portfolio outcomes were heavily influenced by a small number of concentrated positions. Model behavior was characterized by high portfolio concentration, persistence in ticker-level theses, and asymmetric downside exposure. The model frequently re-entered tickers with prior realized losses and relied on event-driven outcomes, contributing disproportionately to overall portfolio drawdowns.结果发现，组合表现极度依赖少数几只重仓股。模型特征明显：持仓集中，对特定股票执着，且下行风险不对称。模型常买回之前亏损的股票，且依赖事件驱动。这导致组合回撤严重。
Rather than optimizing performance or assessing predictive skill, this evaluation examines how an LLM allocates capital, manages risk, and exhibits trading behavior under conditions of limited capital and elevated risk within the micro-cap equity universe, providing insight into structural decision-making tendencies relevant to future LLM-based financial systems.本评估不为优化业绩，也不测预测能力。只看在微盘股这种高风险、资金有限的环境下，LLM 如何调配资金、管理风险、表现行为。为未来的 LLM 金融系统提供参考。
While machine learning has been utilized in finance for decades, systematic research into the behavior of LLMs as autonomous decision-makers in financial contexts remains in its early stages.机器学习在金融界用了几十年。但 LLM 作为自主决策者，研究尚在起步阶段。
Finance serves as a valuable frontier for LLM research due to the complexity and constant evolution of equity markets. In these environments, LLMs can quantitatively evaluate many factors that affect decision-making and behavior under uncertainty.股市复杂且不断演变，是 LLM 研究的绝佳前沿。在这种环境下，LLM 能定量评估多种因素，观察不确定性下的决策与行为。
This evaluation focuses on U.S. listed micro-cap equities. Micro-cap stocks are characterized by higher volatility, lower liquidity, and limited analyst coverage, resulting in greater informational asymmetry. Trading in this environment provides a setting in which information processing, narrative interpretation, and risk management decisions have amplified consequences.本研究聚焦美股微盘股。这类股票波动大、流动性差、分析师覆盖少，信息不对称严重。在这种环境里交易，信息处理、叙事解读和风险管理的后果会被放大。
By constraining ChatGPT to operate exclusively within this environment, this evaluation seeks to observe how an LLM behaves when informational efficiency is weaker and downside risk is elevated. Rather than prioritizing performance or optimization, this evaluation seeks to identify decision-making behavior, portfolio construction tendencies, and observed failure modes under conditions of limited capital and heightened risk.将 ChatGPT 限制在此环境，观察信息效率低下、下行风险高时，它如何表现。不求业绩，只求找出其决策模式、持仓倾向和失效逻辑。
This study seeks to answer: "When ChatGPT is placed in a forward-only trading environment as a portfolio-managing agent given limited starting capital, what behavior patterns emerge?"本研究旨在回答：当 ChatGPT 被置于单向模拟交易环境，且初始资金有限时，会形成什么样的行为模式？
ChatGPT functioned as a decision-maker within the experiment. Trading style, risk management, and position sizing were determined entirely by the model and human input was constrained to manually inputting trades and prompting. The study was strictly confined to stocks in the micro-cap sector in U.S. listed exhanges with limited starting capital ($100). The experiment timeframe was approximately 6 calendar months.ChatGPT 是决策者。交易风格、风控、仓位全由模型定。人只负责手动输入指令。研究严格限于美股微盘股，初始资金 100 美元。实验历时约 6 个月。
The study contributes detailed documentation of decision behavior, execution results, and observed failure modes in the limited information and high volatility of micro-cap equities. The study documents decision behavior within a fixed experimental setup and is not intended as a general assessment of LLM trading ability or as a deployable trading system.本研究记录了模型在微盘股高波动、信息匮乏环境下的决策行为、执行结果及失效模式。仅作固定实验设置下的行为记录，非通用交易能力评价，亦非可直接部署的交易系统。
Portfolio and trade log data were updated manually after each NYSE trading day using a standardized processing script, which generated a structured daily input summary (see Appendix C.3). This summary was provided to the language model as the sole input for decision-making. If trade actions were requested, they were executed on the subsequent trading day. All market data were restricted to only regular trading hours; no pre-market or after-hours data were collected or used.每日收盘后，用脚本汇总组合数据，生成摘要（见附录 C.3）。模型只看这个摘要做决策。若有买卖指令，次日执行。仅限常规交易时段，盘前盘后数据不用。
Human involvement was strictly limited to data entry and trade execution. No discretionary overrides or optimizations were applied to model-generated decisions.人只管录入数据和执行交易。模型出的决策，不干预，不优化。
On a limited number of occasions, daily updates could not be performed following market close. In these cases, the missed update was processed using only past data on that market day. The model was explicitly constrained to rely solely on the provided input and was prohibited from accessing external or future information.偶尔错过收盘更新，补录时仅用当日历史数据。模型严禁访问外部或未来信息。
A weekly research cycle was conducted on Fridays using a dedicated deep research prompt (see Appendix C.2) and the "Deep Research" feature was used. When using the "Deep Research" mode, the model will ask clarifying questions. When the model asked for trading guidance, no judgement was given; however, questions regarding rules and constraints were always answered accurately. Any trade actions proposed outside this framework on Fridays were deferred pending inclusion in the weekly research output. The resulting report was archived, and all trade actions outlined were executed during the subsequent trading week.每周五用“深度研究”提示词进行分析。模型若问交易建议，不予评判；规则类问题则如实回答。周五提出的交易，统一留到下周执行。报告归档，不影响决策逻辑。
This dataset includes overall daily portfolio data for equity and cash, and also includes individual ticker data for each given data. Trade logs were kept in the event of both buying and selling of securities. See Appendix A.0.2 for detailed schema for both CSV files. Raw analytical reports generated during execution were archived in PDF format. Associated textual summaries were also recorded; however, neither the raw reports nor the summaries were incorporated into the analyses presented in this report.含每日组合权益、现金及个股数据。买卖均有记录。文件架构见附录 A.0.2。执行期间生成的原始分析报告存为 PDF，文本总结也留存，但本报告分析不直接引用这些文本。
All benchmark and portfolio data are recorded at daily frequency, with values reflecting end-of-day observations.基准与组合数据均为日频，取收盘值。
The experiment covers the period from June 27, 2025 to December 26, 2025, with all portfolio and benchmark data recorded within this timeframe.实验时间：2025 年 6 月 27 日至 2025 年 12 月 26 日。
-
(Figure 1–2) Equity series + equity-curve metrics (MDD, largest run) → Appendix A.1(图 1–2) 权益序列 + 曲线指标 → 附录 A.1
-
(Figure 6) Holding-period distribution of FIFO lot exits → Appendix A.2 and Appendix D.2(图 6) FIFO 批次持仓分布 → 附录 A.2 和 D.2
-
(Figures 3–4) Position-level (“Pure PnL”) results → Appendix A.3 and Appendix D.3–D.4(图 3–4) 个股盈亏结果 → 附录 A.3 和 D.3–D.4
-
(Figure 5) Episode metrics + PCR → Appendix A.13–A.15 and Appendix D.7(图 5) 交易片段指标 + PCR → 附录 A.13–A.15 和 D.7
-
(Figure 7) Total logged holding days by ticker (count of unique trading dates with Shares > 0 from Daily Updates.csv) → Appendix A.16 (Total Logged Holding Days by Ticker definition) → Appendix D.9 (Total Logged Holding Days by Ticker table)(图 7) 个股累计持仓天数 → 附录 A.16 和 D.9
-
(Figure 8) Repeated buy-side exposure per ticker (count of buy transactions per ticker from Trade Log.csv, plotted as Buy_Entries - 1) → Appendix A.17 (Buy-Side Entries and Re-Entries definition) → Appendix D.10 (Buy Entries and Re-Entries by Ticker table)(图 8) 个股重复买入次数 → 附录 A.17 和 D.10
This study employs a forward-looking, rule-based observational design with quasi-experimental controls.采用前瞻性、规则驱动的观察法，辅以准实验控制。
The large language model ChatGPT was used as a decision-making engine for the portfolio. The model was tasked with generating daily and weekly trade decisions based exclusively on structured summaries of portfolio state and market data.ChatGPT 作为决策引擎。仅基于结构化的组合状态和市场数据摘要，生成每日及每周交易决策。
The model was restricted to purchasing equities within the microcapitalization universe (market capitalization <= $300 million) on U.S listed stock exchanges. This constraint was imposed to evaluate model behavior in securities characterized by limited institutional coverage and reduced analyst attention.仅限市值 3 亿美元以下的微盘股。机构覆盖少，分析师关注低。
Given these conditions, the model’s reasoning was expected to rely primarily on publicly available disclosures, such as company press releases, and on information typically discussed in retail-focused analyses. This design choice allowed observation of the model’s decision-making processes in environments with sparse formal coverage and higher informational asymmetry.模型推理主要依赖公开披露（如新闻稿）和零售分析。此设计旨在观察模型在信息稀疏、不对称环境下的决策过程。
Market data used for portfolio calculations, metrics, and summaries were sourced from the Python library "yfinance" and restricted to end-of-day observations during regular trading hours. These data were processed into standardized daily input summaries reflecting historical price information, portfolio holdings, and cash balances.市场数据来自 Python 库 yfinance，限收盘价。加工为历史价格、持仓和现金余额的每日摘要。
Although the research process permitted consultation of publicly available web sources for contextual analysis, the language model did not have direct access to external websites, raw market data feeds, or real-time information at decision time. Instead, the model operated exclusively on the structured summaries provided as input.模型无直接联网权限，无实时数据源。仅处理输入的结构化摘要。
Weekly research reports and output summaries generated during the study were archived for documentation and analysis purposes. These materials were not incorporated into subsequent model inputs and did not influence future decision-making. Textual reports were not analyzed or used for the conclusions stated in this study.周报归档备查，不作为后续输入，不影响决策。文本报告不参与本研究结论。
All information supplied to the model was limited to data available as of the close of the relevant trading day. No future market data, post-close information, or subsequent outcomes were included in any model input.所有输入限当日收盘前数据。无未来信息。
NOTE: Trade-level statistics were computed at the FIFO lot level, with partial exits treated as independent realized lots rather than as distinct position entries.注：交易统计按 FIFO 批次计算。部分卖出视为独立批次。
Multiple controls were implemented to mitigate common sources of bias in trading studies. To prevent lookahead bias, all model decisions were generated using only information available prior to trade execution, and all trades were executed on a forward-only basis.设多重控制防偏差。严禁“偷看”未来。决策仅用执行前信息，所有交易前瞻执行。
Human involvement was strictly limited to data entry and execution of model-generated instructions. No discretionary overrides, trade filtering, or post hoc optimizations were applied at any point during the experimental period.人只管录入和执行。不干预，不筛选，不事后优化。
On occasions when daily data updates could not be performed immediately following market close, missed updates were processed using only information available as of that trading day. The model was explicitly constrained to rely solely on the provided historical inputs, ensuring that delayed data entry did not introduce access to future information.数据录入延迟时，仅用当日历史数据，确保不引入未来信息。
Although prompt templates evolved over the course of the evaluation, all changes were limited to clarifying existing rules and improving the consistency and precision of report formatting. No changes were made to decision logic, constraints, or trade selection criteria.提示词模板虽有微调，仅为澄清规则和统一格式。决策逻辑、约束条件和选股标准未变。
Figure 1. Portfolio equity versus benchmarks (normalized to $100) over time.图 1. 投资组合权益与基准对比（归一化至 100 美元）。
Benchmarks such as the Russell 2000 and the S&P 500 were included to provide contextual reference for broad market conditions during the experimental period rather than to evaluate relative performance. The benchmarks serve to anchor the observed equity trajectory within market regimes; however, all conclusions in this evaluation are derived from portfolio behavior, decision patterns, and realized outcomes, and are not contingent on benchmark performance comparisons.罗素 2000 和标普 500 仅作市场环境参考，非业绩对比。结论基于模型行为和决策模式，而非基准表现。
As shown in Figure 1, portfolio equity declined substantially relative to both the Russell 2000 and the S&P 500 over the experimental period.如图 1，实验期内组合权益相对于基准大幅下跌。
Figure 2. Portfolio equity with max drawdown percentage (red) and largest run (green).图 2. 组合权益、最大回撤（红）与最大涨幅（绿）。
Figure 2 highlights the largest positive equity movement and the maximum drawdown observed during the experimental period. The largest run occurred between November 13, 2025 and November 18, 2025, during which portfolio equity increased by 21.51%. The maximum drawdown reached -50.33%, corresponding to an equity value of $67.10 on November 6, 2025.图 2 标出最大涨幅和回撤。2025 年 11 月 13 日至 18 日期间，权益涨 21.51%。11 月 6 日最大回撤达 -50.33%，权益降至 67.10 美元。
Using FIFO lot-level reconstruction, 46 realized lot exits were observed. Exactly 50% of lot exits were profitable; however, average losses exceeded average gains (-3.83 vs +3.01), producing a profit factor of 0.82 and a negative per-lot expectancy of -0.41. Median outcomes showed the opposite pattern, indicating that overall underperformance was driven by a small number of large losses rather than uniformly poor trade selection.FIFO 批次重构显示，共 46 次平仓。胜率 50%。但亏损平均值（-3.83）大于盈利平均值（+3.01），盈亏比 0.82，期望值为负（-0.41）。中位数表现相反，说明整体亏损源于少数大额亏损，而非选股普遍不行。
Full Individual Trade Table found in Appendix D.2.完整交易表见附录 D.2。
Figure 3. Realized PnL (USD) by ticker.图 3. 个股实现盈亏（美元）。
Figure 3 shows 10 of the 22 tickers the model bought within the experimental period generated profits. Profits among tickers generally had concentrated profits; with the exception of ATYR, losses were less concentrated.图 3 显示模型买入的 22 只股票中，10 只盈利。盈利分布集中。除 ATYR 外，亏损相对分散。
Figure 4. Top realized PnL (USD) ticker wins vs. losses.图 4. 盈亏金额前几名个股对比。
As shown in Figure 4, realized losses were larger in magnitude than realized gains. The most significant downside outcome was attributable to ATYR, indicating that overall portfolio performance was strongly influenced by a small number of adverse position-level outcomes.如图 4，亏损额大于盈利额。ATYR 亏损最重，说明少数极端个股决定了组合表现。
On average, the portfolio held 3.1 tickers per trading day, indicating a high degree of concentration throughout the experimental period (Appendix D.8).组合平均每日持仓 3.1 只，高度集中（附录 D.8）。
Figure 5. Peak Capture Ratio (exit PnL divided by peak unrealized PnL) plotted against peak unrealized profit for valid trade episodes.图 5. 有效交易片段的峰值捕获率（已实现盈亏 / 峰值浮盈）。
Four episodes had null Peak Capture Ratios due to the peak recorded PnL value being below zero. These tickers were left out of the graph.4 个片段因峰值浮盈为负，剔除。
Two other trade episodes were also excluded due to methodological reasons. The ATYR episode produced a valid capture ratio; however, realized outcomes were materially affected by execution constraints following a large overnight price gap, requiring a manual exit at the opening price and preventing normal stop-loss execution. As a result, the observed ratio reflects execution limitations rather than exit timing behavior. The FBIO episode, by contrast, represents a structural outlier in which a small unrealized peak preceded a large realized loss, producing an extreme negative capture ratio driven by denominator instability. FBIO is therefore excluded for interpretability.另剔除两例。ATYR 因隔夜大缺口被迫开盘平仓，捕获率反映的是执行限制而非离场时机。FBIO 因浮盈峰值极小导致分母不稳定，产生极端负值，剔除以保可读性。
The definition of a trade episode is provided in Appendix A.13. The complete trade-episode table is available in Appendix D.7.交易片段定义见附录 A.13，完整表见附录 D.7。
Figure 5 indicates that, for the remaining episodes, exits often captured a meaningful fraction of peak unrealized PnL. Peak Capture Ratios varied substantially across episodes, with no clear relationship between peak PnL magnitude and exit timing behavior.图 5 显示，离场时通常能捕获相当一部分浮盈。捕获率差异大，与浮盈规模无明显关联。
Although trade execution occurred at the FIFO lot level, aggregation of realized outcomes at the position level reveals that only 10 of 22 tickers generated positive total PnL. Average losses exceeded average gains in magnitude, and the largest single position loss dominated overall results, consistent with exposure to binary, event-driven return dynamics.22 只股票仅 10 只盈利。平均亏损大于平均盈利，单只股票亏损主导结果，符合事件驱动型交易特征。
The distribution of FIFO lot-level outcomes (reported in Trade-Level Analysis) aligns with the concentration patterns observed at the ticker level, indicating that portfolio-level concentration emerged from a small number of large-magnitude realized exits rather than from uniformly poor trade execution.FIFO 批次结果与个股层面一致，说明集中度风险源于少数大额离场，而非选股能力普遍低下。
Figure 6. Distribution of holding periods across individual closed lots.图 6. 单个平仓批次的持仓时长分布。
Figure 6 shows a strongly right-skewed distribution with a long right tail. The majority of FIFO lot exits occurred within approximately 10–20 days, while a small number of lots were held for substantially longer durations, including one holding period exceeding 70 days calendar days.右偏分布，长尾特征明显。多数在 10-20 天内平仓，少数持仓超 70 天。
Figure 7. Total individual ticker holding duration during trading days.图 7. 个股累计交易日持仓时长。
Figure 7 shows that cumulative holding time was concentrated in a small number of tickers, with ABEO and MIST accounting for the largest total time-in-portfolio exposure. Despite their extended holding durations, these positions produced divergent realized outcomes: ABEO ranked fifth in total realized PnL, while MIST ranked eighth from the bottom. This contrast indicates that prolonged holding time alone was not a reliable determinant of portfolio-level performance.ABEO 和 MIST 持仓最久。但 ABEO 盈利排名第五，MIST 倒数第八。说明持仓久不代表业绩好。
Note: Figure 6 reports FIFO lot holding durations in calendar days to capture individual lot persistence, while Figure 7 measures cumulative holding time in trading days to reflect overall market exposure.注：图 6 按自然日算批次持仓，图 7 按交易日算个股累计持仓。
Instead, individual ticker outcomes were more impactful for overall performance than holding duration.个股盈亏对业绩影响远大于持仓时长。
Figure 8. Number of repeated buy-side trade entries per ticker.图 8. 个股重复买入次数。
Seven of the 22 tickers were purchased on multiple occasions. Notably, the three tickers with the lowest realized PnL (see Figure 4) were all subject to repeated buy-side entries. In contrast, of the top three highest PnL tickers in Figure 3, only ALDX was repurchased more than once.22 只股票中 7 只被多次买入。亏损最重的三只股票均被多次买入。而盈利最高的三只中，仅 ALDX 被重复买入。
This pattern suggests persistence in position-level theses, with the model exhibiting limited responsiveness to realized performance when determining whether to re-enter previously traded securities.这表明模型对个股逻辑有执念，对亏损后的再入场缺乏应变。
This section examines how selected aggregate performance metrics change when the ATYR positions, the largest realized loss during the experimental period, is omitted. The purpose is to assess the sensitivity of portfolio-level summaries to single-position tail events rather than to reinterpret the primary results.考察剔除最大亏损源 ATYR 后的指标变化。旨在评估单点风险对组合的影响。
Average loss slightly improved from -$3.63 to -$2.22. Expectancy and profit factor, however, had major reversals. Profit factor improved from 0.82 to 1.52, and expectancy went from -0.41 to 0.49.平均亏损从 -3.63 改善至 -2.22。盈亏比从 0.82 升至 1.52，期望值从 -0.41 升至 0.49。
Notably, these reversals occur without meaningful changes in win rate or median lot outcomes, suggesting that aggregate underperformance is attributable to tail-risk exposure rather than decision failure.胜率和中位数没变，说明整体表现不佳源于尾部风险，而非决策失误。
Metrics and tables discussed in this section are reported in Appendix D.12.相关指标见附录 D.12。
Human input was needed in the execution loop for inputting trades and prompting for each trading day.每日需人录入数据并提示模型。
The model was only allowed to buy U.S. listed tickers with market capitalizations equal to or less than 300M. If a held ticker's capitalization became greater than 300M, the model could not buy any more shares.仅限市值 3 亿以下美股。若持仓股涨超 3 亿，则不能加仓。
All data was calculated based on end-of-day trading data only.所有计算仅基于收盘数据。
The trading simulation did not incorporate transaction costs such as commissions or bid–ask spread effects.不计佣金和买卖价差。
Financial derivatives were strictly not allowed.严禁使用金融衍生品。
Throughout the experiment, the portfolio routinely consisted of 2-3 concentrated tickers. The overall portfolio was particularly sensitive to individual ticker factors.组合常维持 2-3 只重仓股，对单只股票波动极其敏感。
As shown in Figure 4 and Figure 7, FBIO and IINN were among the largest contributors to realized losses. Both tickers had relatively high trade frequencies, with FBIO having four buy-side trades and IINN receiving three.FBIO 和 IINN 亏损最重，但交易频率高，分别被买入 4 次和 3 次。
Trade selection was characterized by exposure to event-driven catalysts (e.g., regulatory announcements), associated with large positive or negative outcomes. Comparatively limited exposure was observed in lower-volatility or incremental-return investing.偏好事件驱动（如监管公告），博取大涨大跌，缺乏稳健的增量收益策略。
Throughout the experiment, the model's portfolio consisted of three tickers for trading days on average. The average cost basis for a ticker position was $25.28 (25% of starting capital). This indicates the model preferred high concentration positions with limited portfolio diversity. This allocation pattern reflects a preference for highly concentrated positions and limited diversification, resulting in substantial exposure to individual ticker-level movements.平均每日持仓 3 只。平均单只成本占初始资金 25%。模型偏好高集中度，缺乏多样化，导致单只股票波动对组合影响巨大。
The model’s performance was characterized by pronounced loss asymmetry. As shown in Figure 2, the portfolio reached a maximum drawdown of approximately 50% over the experimental period. The largest losing ticker (ATYR) generated losses more than double the gains produced by the largest winning ticker (SLS). On the day of ATYR’s sharp decline, the portfolio experienced an approximately 40% single-day equity drop. Loss asymmetry was driven primarily by this single event, as other losing positions were comparatively small and did not materially affect portfolio-level outcomes.亏损严重不对称。最大回撤 50%。ATYR 亏损是最大赢家 SLS 盈利的两倍多。ATYR 单日大跌导致组合单日回撤 40%。亏损主要由单次极端事件驱动。
According to Figure 8, the model bought approximately 32% of tickers entered during the experimental period more than once (7 of 22). As shown in Figure 3, five of these repeatedly traded tickers generated negative cumulative realized PnL across all engagements. In contrast, the two of the highest-PnL tickers in Figure 3 were each purchased only once, with the exception of ALDX, during the experimental period. This pattern indicates persistence in position-level exposure, with re-entry occurring despite unfavorable prior outcomes rather than avoidance following realized losses.32% 的股票被多次买入。其中 5 只累计亏损。而盈利最高的股票大多只买了一次。模型对旧逻辑有执念，不因亏损而回避。
Figure 6 shows the majority of exits occurred within approximately 10–20 calendar days, while a small number of lots were held for substantially longer durations, including one holding period exceeding 70 total days. Holding duration, however, did not correlate with substantially higher or lower realized PnL. As shown in Figure 7, the two tickers with the highest cumulative holding time, MIST and ABEO, did not rank among the top or bottom three tickers by total realized PnL in Figure 4.多数在 10-20 天离场，少数超 70 天。时长与盈亏无相关性。持仓最久的 MIST 和 ABEO 业绩平平。
These observations are purely qualitative and are not analyzed quantitatively in the evaluation due to limitations of collected data and research scope.以下为定性观察，数据有限，不做定量分析。
During execution, stop-loss levels were often adjusted upward prior to anticipated catalyst events. However, large adverse price moves occurred outside regular trading hours, rendering these stop-loss levels ineffective. The model appeared to persist in this behavior even after the substantial equity loss associated with the ATYR outcome.止损位常在事件前上调。但大波动多发生在盘后，止损失效。ATYR 事件后，模型仍未改变此行为。
The model would occasionally hallucinate portfolio details or explicitly state rationale that was not grounded in the experimental setup during Deep Research reports. For example, in one weekly report the model referenced the absence of using hedging instruments, despite the use of derivatives and short positions being clearly prohibited throughout the experiment. This specific inconsistency reflects narrative generation rather than a change in execution logic. A representative excerpt is provided in Appendix B.模型偶尔会产生幻觉，在周报中声称未使用对冲工具，尽管实验全程禁止衍生品。这反映了叙事生成与实际逻辑的脱节。
Due to the limited experimental period, the data may not be representative of the model's behavior across different market regimes. The evaluation should only be analyzed in the context of micro-cap equities; LLM behavior may vary widely in different market capitalizations. The results analyzed are based on a single experimental run and do not capture variability across repeated runs or alternative random initial conditions. This study is descriptive in nature, documenting observed decision-making behavior when an LLM is placed in a constrained capital allocation role.周期短，样本代表性不足。仅限微盘股，其他市场表现可能不同。单次实验，未考虑随机变量。研究仅为描述性记录。
The experimental setup relied on interactive use of the publicly available ChatGPT interface, introducing sources of variability that could not be fully controlled. Prompt templates evolved modestly to clarify existing constraints, newer model versions were adopted as they became available, and generation parameters such as temperature were not explicitly fixed.交互式使用 ChatGPT 引入了不可控变量。提示词模板微调，模型版本更新，温度参数未固定。
The evaluation focuses on observable decision behaviors under consistent informational and procedural constraints rather than on comparisons across specific model configurations.评估聚焦于约束条件下的决策行为，而非模型配置对比。
Across the experimental period, portfolio equity outcomes were dominated by a small number of high-impact trades. High position concentration amplified exposure to individual ticker outcomes, with a single adverse position exerting a disproportionate influence on overall portfolio balance. Trading behavior exhibited persistence in position-level theses, as the model re-entered tickers despite prior exits, including cases with realized losses. Tickers subject to repeated buy-side entries accounted for the largest cumulative equity losses. In addition, the model held several positions for extended durations despite ultimately contributing minimally to overall portfolio performance, indicating that holding period length alone was not a reliable determinant of realized outcomes. Exit outcomes frequently occurred after a substantial portion of unrealized gains had accrued, though full capture of peak PnL was uncommon.组合表现由少数高影响交易决定。高集中度放大了单只股票风险。模型对持仓逻辑执着，即便亏损也重复买入。重复买入的股票反而贡献了最大的亏损。持仓时长与业绩无关。离场时往往已回吐大部分浮盈。
Taken together, these results suggest that, when placed in a capital allocation role, an LLM exhibits decision-making patterns resembling high-conviction, thesis-driven discretionary trading. Portfolio outcomes were shaped less by incremental trade-level performance and more by concentration, persistence in position-level narratives, and asymmetric downside exposure.总而言之，LLM 在资金调配中表现出高信念、叙事驱动的自由交易风格。结果更多由集中度、叙事执念和不对称下行风险决定，而非交易层面的优化。
Future evaluations could reduce uncontrolled variability by replacing interactive human prompting with algorithmic prompting and fixed generation parameters. Programmatic control over sampling settings, such as temperature, would allow differences in experimental outcomes to be attributed to stochastic variation rather than procedural differences. This modification would support more rigorous comparison across runs and across models, while preserving the forward-only execution constraints of the study.未来可改人工提示为算法提示，固定生成参数。通过程序控制温度等采样设置，区分随机偏差与逻辑偏差，实现更严谨的对比。
Future comparisons across sectors in the stock market could reveal model patterns given different environments. Factors such as risk management, report confidence, and behavioral changes following market events could be used to evaluate how decision-making differs across market scenarios.对比不同行业表现，观察模型在不同市场环境下的风控与行为变化。
Future work could examine differences in decision-making behavior across large language models under identical experimental conditions. Comparing model behavior could highlight differences in risk management, concentration, persistence, and responses to realized outcomes when acting as portfolio decision-makers.在相同条件下对比不同 LLM，突出其在风控、集中度、执念及亏损响应上的差异。
Evaluations conducted under identical constraints, with the exception of increased starting capital, would likely produce different risk management behaviors. Future work could analyze how starting capital influences position sizing, diversification, and overall portfolio performance.改变初始资金规模，分析其对仓位管理、多样化及整体业绩的影响。
By analyzing the collected reports over the course of the experiment, future work could examine the relationship between expressed sentiment and realized portfolio performance over time. This analysis could also investigate divergences between generated report language and observed model behavior, including cases where narrative confidence did not align with portfolio actions that followed. In addition, the frequency and nature of model hallucinations, such as false rule assumptions or incorrect portfolio state descriptions, could be systematically documented and analyzed.分析报告情感与业绩的关系。调查叙事自信与实际操作的偏离。系统记录模型幻觉（如错误规则假设或组合状态描述错误）的频率与性质。
Portfolio monitoring and recordkeeping in this study used a blended data collection process. Two primary artifacts were maintained:研究采用混合收集流程，维护两类记录：
Trade Log.csv, containing discrete buy and sell events including dates, shares, prices, and cost basis fieldsTrade Log.csv：含买卖日期、股数、价格、成本基础。Daily Updates.csv, containing end-of-day position snapshots including shares held and per-day PnL fields by tickerDaily Updates.csv：含收盘持仓快照、个股每日盈亏。
These sources were produced through an operational workflow combining manual updates and a standardized processing script. They were not generated as a single unified, transaction-perfect ledger. As a result, the analyses in this appendix should be interpreted as reconstructions derived from blended records rather than as broker-grade, audit-ready accounting statements.来源结合了人工更新与脚本处理，非单一、完美的事务账本。本附录分析基于重构数据，非审计级会计报表。
The analyses in this report rely on the following source-file column schemas.本报告分析，基于以下源文件列结构。
Daily Updates.csv columns:Daily Updates.csv 各列：
- Date日期 (Date)
- Ticker代码 (Ticker)
- Shares股数 (Shares)
- Buy Price买入价 (Buy Price)
- Cost Basis成本基准 (Cost Basis)
- Stop Loss止损价 (Stop Loss)
- Current Price当前价 (Current Price)
- Total Value总值 (Total Value)
- PnL盈亏 (PnL)
- Action动作 (Action)
- Cash Balance现金余额 (Cash Balance)
- Total Equity总权益 (Total Equity)
Trade Log.csv columns:Trade Log.csv 各列：
- Date
- Ticker
- Shares Bought买入股数 (Shares Bought)
- Buy Price
- Cost Basis
- PnL
- Reason原因 (Reason)
- Shares Sold卖出股数 (Shares Sold)
- Sell Price卖出价 (Sell Price)
Metrics in Appendix A are computed using reconstruction strategies designed to produce consistent, interpretable summaries of behavior:附录 A 指标，由重构策略计算。旨在总结行为，求其一致与可读：
- FIFO lot-level realized exits are reconstructed from
Trade Log.csvby matching sell transactions to prior buy transactions under a first-in, first-out conventionFIFO 批次实现盈亏，从 Trade Log.csv 重构。卖出交易与先前的买入交易匹配，依先进先出原则。 - Position-level (Pure PnL) statistics are reconstructed by aggregating reconstructed FIFO lot exits to one row per ticker持仓级（纯盈亏）统计，由重构后的 FIFO 批次汇总。每只代码一行。
- Episode-level statistics (including Peak Capture Ratio) are reconstructed from
Daily Updates.csvby identifying continuous holding intervals from theSharesfield and summarizing the recorded dailyPnLseries周期级统计（含峰值捕获率），从 Daily Updates.csv 重构。依股数字段找连续持有区间，汇总每日盈亏序列。 - Equity-curve statistics (including Maximum Drawdown and Largest Run) are reconstructed from
Daily Updates.csvusing rows whereTicker == "TOTAL"and theTotal Equityfield, with a baseline row inserted权益曲线统计（含最大回撤与最大涨幅），从 Daily Updates.csv 重构。取代码为 "TOTAL" 的行，用总权益字段，插入基准行。
These reconstructions approximate realized and unrealized dynamics for analytical purposes. They are not intended to be treated as definitive accounting of tax lots, brokerage fills, corporate actions, intraday execution effects, or all sources of slippage.重构仅为分析。非税务批次、券商成交、公司行动、日内执行或滑点来源的最终账目。
Because the underlying dataset is blended and reconstructed, the metrics are intended to support behavioral characterization (for example: concentration, persistence, re-entry frequency, and exit timing tendencies) rather than precise performance attribution.数据集经混合重构。指标用于刻画行为（如集中度、持久性、重入频率、退出时机），而非精确的业绩归因。
Where reconstruction choices matter, this appendix defines the conventions explicitly (FIFO lot matching, calendar-day holding duration, episode segmentation, and equity-curve segmentation) so that results can be reproduced under the same assumptions.重构逻辑见附录（FIFO 匹配、日历日持有、周期切分、权益曲线切分）。依此假设，结果可复现。
Maximum drawdown and largest run calculations are computed from the portfolio equity time series constructed as follows:最大回撤与最大涨幅，由组合权益时间序列计算：
- Filter
Daily Updates.csvto rows whereTicker == "TOTAL"过滤 Daily Updates.csv，取代码为 "TOTAL" 的行。 - Parse
Dateas a datetime将日期解析为 datetime。 - Coerce
Total Equityto numeric将总权益强制转为数值。 - Prepend a baseline row:
- Date = 2025-06-27日期 = 2025-06-27
- Total Equity = 100.0总权益 = 100.0
- Concatenate, sort by Date, and drop duplicate Dates keeping the last record for each Date合并、按日期排序。去重，保留各日期最后一条记录。
This produces a single Date-sorted equity series used for all equity-curve metrics.生成单一按日期排序的权益序列，用于所有权益曲线指标。
The running maximum at time t is defined as the maximum observed portfolio equity from the start of the series through time t.t 时刻的滚动最大值，为序列开始至 t 时刻观察到的最大组合权益。
In code, this is computed as:代码计算如下：
- Running Max = cumulative maximum of
Total Equity滚动最大值 = 总权益的累计最大值
Drawdown percentage at time t is computed as:t 时刻回撤百分比计算如下：
- Drawdown % = (
Total Equity/Running Max- 1.0) × 100.0回撤 % = (总权益 / 滚动最大值 - 1.0) × 100.0
This produces values of 0.0 at new equity highs and negative values when equity is below the prior peak.创新高时为 0.0。权益低于前高时为负值。
Maximum drawdown is the most negative drawdown percentage observed across the full series.最大回撤为全序列中观察到的最小回撤百分比。
In code, the maximum drawdown point is selected by:代码确定最大回撤点：
- finding the row with the minimum value of
Drawdown %找到回撤 % 最小值的行。
Reported maximum drawdown fields:最大回撤字段：
- max_drawdown_date: Date of the minimum
Drawdown %rowmax_drawdown_date：最小回撤 % 发生的日期。 - max_drawdown_equity:
Total Equityon that Datemax_drawdown_equity：该日期的总权益。 - max_drawdown_pct:
Drawdown %on that Datemax_drawdown_pct：该日期的回撤 %。
Largest Run is defined as the largest percentage increase from a local minimum to a subsequent peak, using the segmentation logic in find_largest_gain(...).最大涨幅为从局部最低点到后续峰值的最大涨幅百分比。逻辑见 find_largest_gain(...)。
Algorithm summary:算法摘要：
- Initialize the first observation as the current local minimum and current peak初始化第一个观察值为当前局部最低点与当前峰值。
- If a new higher equity value occurs, update the current peak若出现更高权益，更新当前峰值。
- If a decline occurs (current equity < current peak), compute the completed run gain:
- gain = (peak_val - min_val) / min_val × 100.0涨幅 = (峰值 - 最低值) / 最低值 × 100.0
- update the best run if this gain exceeds the prior best若涨幅超前值，更新最佳涨幅。
- reset local minimum and peak to the current observation将局部最低点与峰值重置为当前观察值。
- After reaching the end of the series, also evaluate the final run segment序列结束时，评估最后一段。
Reported largest run fields:最大涨幅字段：
- largest_run_start: Date of the local minimum for the best runlargest_run_start：最佳涨幅的局部最低点日期。
- largest_run_end: Date of the subsequent peak for the best runlargest_run_end：最佳涨幅的后续峰值日期。
- largest_run_gain_pct: percent gain for the best runlargest_run_gain_pct：最佳涨幅百分比。
Interpretation note:解读注：
- This is a segment-based definition determined by the reset-on-decline rule above, rather than a global search over all possible minimum-to-maximum intervals.此为基于上述“回撤重置规则”的片段定义，非全局所有最低点到最高点区间的搜索。
A FIFO lot is defined as a discrete group of shares created by a single buy transaction in Trade Log.csv.FIFO 批次，指 Trade Log.csv 中由单笔买入交易创建的一组离散股票。
In the reconstruction code, each buy transaction creates one open lot with:重构代码中，每笔买入创建一笔开仓批次：
- Entry_Date: the transaction
DateEntry_Date：交易日期。 - Shares:
Shares BoughtShares：买入股数。 - Entry_Price:
Cost Basis / Shares Bought(per-share)Entry_Price：成本基准 / 买入股数（每股价格）。
Lots are ordered chronologically by Entry_Date and are matched to sells using a first-in, first-out accounting convention.批次按 Entry_Date 排序，依先进先出原则与卖出匹配。
A FIFO lot-level realized exit occurs when some or all shares from one or more open FIFO lots are closed via a sell transaction.FIFO 批次实现退出，指通过卖出交易平掉部分或全部开仓批次。
In the reconstruction code, when a sell occurs:重构代码中，卖出时：
- The sell quantity (
Shares Sold) is allocated to the oldest open lots first卖出数量（Shares Sold）优先分配给最旧的开仓批次。 - Partial lot closures are permitted允许部分平仓。
- Each partial closure is recorded as a distinct realized exit record每笔部分平仓记录为独立的实现退出记录。
For each realized exit record, the following fields are recorded:每条实现退出记录包含以下字段：
- Ticker: security identifierTicker：证券标识。
- Entry_Date: date the lot was openedEntry_Date：开仓日期。
- Exit_Date: date the shares were soldExit_Date：卖出日期。
- Shares: number of shares closed from the lotShares：从批次平掉的股数。
- Entry_Price: per-share entry price of the lotEntry_Price：批次每股入场价。
- Exit_Price: per-share exit price (taken from
Sell Price)Exit_Price：每股卖出价（取自 Sell Price）。 - PnL: realized profit or loss in USD, computed as
Shares × (Exit_Price - Entry_Price)PnL：以美元计的实现盈亏，计算为 Shares × (Exit_Price - Entry_Price)。 - Holding_Days: calendar days between Entry_Date and Exit_Date, computed as
(Exit_Date - Entry_Date).daysHolding_Days：Entry_Date 与 Exit_Date 之间的日历天数，计算为 (Exit_Date - Entry_Date).days。
All FIFO lot-level performance metrics are computed from this realized exit table.所有 FIFO 批次业绩指标均基于此实现退出表计算。
Pure PnL aggregates FIFO lot-level realized exits to the ticker (position) level.纯盈亏将 FIFO 批次实现退出汇总至代码（持仓）级。
Each ticker is represented by a single row summarizing all realized FIFO lot exits for that security.每只代码由一行记录，总结该证券所有实现的 FIFO 批次退出。
For each ticker, the following metrics are computed from the FIFO lot exit table:每只代码的指标由 FIFO 批次退出表计算：
-
PnL Sum of realized PnL across all FIFO lot exit records for the tickerPnL：该代码所有 FIFO 批次退出记录的实现盈亏总和。
-
Holding_Days Calendar-day span from the earliest Entry_Date to the latest Exit_Date for the ticker, computed as
(max(Exit_Date) - min(Entry_Date)).daysHolding_Days：从最早 Entry_Date 到最晚 Exit_Date 的日历天数，计算为 (max(Exit_Date) - min(Entry_Date)).days。 -
Avg_Position_Size Computed as
np.average(Shares × Entry_Price, weights=Shares), using the FIFO lot exit records for the ticker Equivalently:sum( Shares_i × (Shares_i × Entry_Price_i) ) / sum(Shares_i)Avg_Position_Size：计算为 np.average(Shares × Entry_Price, weights=Shares)。即 sum(Shares_i × (Shares_i × Entry_Price_i)) / sum(Shares_i)。 -
Num_Lot_Exits Number of FIFO lot exit records for the tickerNum_Lot_Exits：该代码的 FIFO 批次退出记录数。
Pure PnL metrics are used to analyze concentration and ticker-level contribution to realized performance under the stated reconstruction conventions.纯盈亏指标用于分析集中度及各代码对实现业绩的贡献。
For any metric set computed using the reconstructed tables:对任何重构表计算的指标集：
- Wins are observations where
PnL > 0赢：PnL > 0 的观察值。 - Losses are observations where
PnL < 0亏：PnL < 0 的观察值。 - Zero-PnL observations (
PnL == 0) are neither wins nor losses零盈亏（PnL == 0）既非赢也非亏。
The metrics define:指标定义如下：
- Count as the total number of observations, including any zero-PnL observationsCount：观察值总数，含零盈亏。
- Win Rate using Count as the denominatorWin Rate：以 Count 为分母。
Average and median win/loss statistics are computed only over the corresponding subsets (wins or losses).平均与中位数赢/亏统计，仅在相应子集（赢或亏）中计算。
Count is the total number of observations in the relevant input table:Count 为相关输入表中的观察值总数：
- FIFO lot-level metrics: number of FIFO lot exit recordsFIFO 批次级：FIFO 批次退出记录数。
- Position-level metrics: number of ticker rows in the Pure PnL table持仓级：纯盈亏表中的代码行数。
Win Rate is defined as the fraction of observations with positive realized PnL:胜率为实现盈亏为正的观察值占比：
Win Rate = Number of observations with PnL > 0 / Total number of observations胜率 = PnL > 0 的观察值数 / 观察值总数
The denominator includes any zero-PnL observations.分母包含零盈亏。
Average Win is the arithmetic mean of PnL over observations with PnL > 0.平均赢为 PnL > 0 观察值的算术平均值。
Median Win is the median of PnL over observations with PnL > 0.中位数赢为 PnL > 0 观察值的中位数。
Average Loss is the arithmetic mean of PnL over observations with PnL < 0.平均亏为 PnL < 0 观察值的算术平均值。
Loss values are reported as negative numbers.亏损值记为负数。
Median Loss is the median of PnL over observations with PnL < 0.中位数亏为 PnL < 0 观察值的中位数。
Profit Factor is defined as:盈亏比定义为：
Profit Factor = Sum(PnL over wins) / absolute value of Sum(PnL over losses)盈亏比 = 赢利总和 / 亏损总和的绝对值
If Sum(PnL over losses) equals zero, Profit Factor is reported as infinite.若亏损总和为零，盈亏比记为无穷大。
Expectancy represents the expected PnL per observation:期望值为单次观察的预期盈亏：
Expectancy = (Average Win × Win Rate) + (Average Loss × (1 - Win Rate))期望值 = (平均赢 × 胜率) + (平均亏 × (1 - 胜率))
Expectancy is expressed in USD per observation.以美元/观察值计。
Average Holding Days is the arithmetic mean of the relevant holding-duration column:平均持有天数为相关持有时长列的算术平均值：
- FIFO lot-level metrics: mean of lot-level Holding_Days values in calendar daysFIFO 批次级：批次级 Holding_Days 的均值（日历天数）。
- Position-level metrics: mean of ticker-level Holding_Days values持仓级：代码级 Holding_Days 的均值。
Holding days are computed as calendar-day differences and do not adjust for non-trading days.按日历天数计算，不调整非交易日。
A ticker is classified as having repeated exposure if it has more than one FIFO lot exit record during the experimental period.若某代码在实验期内有超过一条 FIFO 批次退出记录，则视为有重复敞口。
This reflects multiple realized lot exits for the same ticker and does not necessarily imply multiple distinct entry episodes.此反映同一代码的多次实现退出，未必意味着多次独立的入场周期。
Average Tickers Held Per Day is computed from Daily Updates.csv (excluding rows where Ticker == "TOTAL") as:由 Daily Updates.csv 计算（排除代码为 "TOTAL" 的行）：
- For each Date, count the number of unique tickers present in the file for that Date对每个日期，统计文件中唯一代码数。
- Take the mean of that daily unique-ticker count across all Dates取所有日期唯一代码数的均值。
Because Daily Updates.csv includes only tickers with Shares > 0, this measure equals the average number of tickers held per trading day.因 Daily Updates.csv 仅含 Shares > 0 的代码，此即日均持仓代码数。
Average Ticker Cost Basis is computed from Daily Updates.csv (excluding rows where Ticker == "TOTAL") as the arithmetic mean of the Cost Basis column across all remaining ticker-day rows.由 Daily Updates.csv 计算（排除代码为 "TOTAL" 的行），为所有剩余代码-日行的成本基准列算术平均值。
This is a ticker-day average based on the daily snapshot file and its row conventions.基于每日快照文件的代码-日平均值。
A trade episode is defined using Daily Updates.csv as a continuous interval where Shares > 0 for a given ticker.交易周期由 Daily Updates.csv 定义，指某代码 Shares > 0 的连续区间。
- An episode begins on the first Date where
Shares > 0following a prior Date whereShares == 0(or no prior record)周期始于 Shares > 0 的首个日期（其前一日 Shares == 0 或无记录）。 - An episode ends on the last Date prior to a return to
Shares == 0周期止于回到 Shares == 0 前的最后一个日期。 - Episodes are identified independently for each ticker各代码独立识别周期。
For each ticker episode, the following are computed from the subset of rows where Shares > 0:对每个代码周期，由 Shares > 0 的行子集计算：
- start_date: first Date of the episodestart_date：周期首日。
- end_date: last Date of the episodeend_date：周期末日。
- peak_pnl: maximum value of the
PnLfield within the episodepeak_pnl：周期内 PnL 字段的最大值。 - exit_pnl: last value of the
PnLfield within the episode (PnL on the final episode day)exit_pnl：周期内 PnL 字段的最后值（周期末日盈亏）。 - duration_days: calendar-day span from start_date to end_date, computed as
(max(Date) - min(Date)).daysduration_days：从 start_date 到 end_date 的日历天数，计算为 (max(Date) - min(Date)).days。
All episode PnL quantities refer to the recorded daily PnL field in Daily Updates.csv.所有周期盈亏量，均指 Daily Updates.csv 中记录的每日盈亏字段。
Peak Capture Ratio is defined as:峰值捕获率定义为：
Peak Capture Ratio = exit_pnl / peak_pnl峰值捕获率 = exit_pnl / peak_pnl
If peak_pnl <= 0, Peak Capture Ratio is set to null.若 peak_pnl <= 0，峰值捕获率设为 null。
This rule is applied to remove undefined or non-informative ratios under the stated episode convention.此规则旨在剔除周期定义下无意义或未定义的比率。
Total Logged Holding Days by Ticker measures how many trading days each ticker had exposure in the portfolio.代码总记录持有天数，衡量各代码在组合中的交易日敞口。
This metric is computed from Daily Updates.csv as:由 Daily Updates.csv 计算：
- Exclude rows where
Ticker == "TOTAL"排除代码为 "TOTAL" 的行。 - Filter to rows where
Shares > 0过滤 Shares > 0 的行。 - Group by
Ticker按代码分组。 - Count the number of unique
Datevalues per ticker统计各代码的唯一日期数。
This counts trading days with exposure (not calendar days).此统计有敞口的交易日，非日历天数。
The resulting series is used to generate Figure 7.所得序列用于生成图 7。
A buy entry is defined as any row in Trade Log.csv where Shares Bought is present (non-null).买入指 Trade Log.csv 中 Shares Bought 字段存在（非空）的行。
(Optionally, treat a buy entry as Shares Bought > 0 if the file can contain zeros.)（若文件含零，可选将 Shares Bought > 0 视为买入。）
Buy Entries per Ticker is computed as:各代码买入次数计算如下：
- Filter
Trade Log.csvto rows whereShares Boughtis non-null (or > 0)过滤 Trade Log.csv，取 Shares Bought 非空（或 > 0）的行。 - Group by
Ticker - Count rows per ticker按代码统计行数。
In code, this corresponds to:代码实现：
buy_entries = trades_df.dropna(subset=["Shares Bought"]).groupby("Ticker").size()buy_entries = trades_df.dropna(subset=["Shares Bought"]).groupby("Ticker").size()
Re-Entries per Ticker measures repeated buy-side exposure after an initial entry.重入次数衡量首次入场后的重复买入敞口。
For each ticker:对每只代码：
- Re_Entries = max(Buy_Entries - 1, 0)Re_Entries = max(Buy_Entries - 1, 0)
This produces:结果为：
- 0 if a ticker was bought once买入一次记 0。
- 1 if a ticker was bought twice买入两次记 1。
- etc.以此类推。
The resulting series is used to generate Figure 8.所得序列用于生成图 8。
Context: Deep Research weekly report. The following excerpts are included to illustrate post hoc narrative framing around constraints and hedging within the experiment’s rule set.背景：深度研究周报。以下摘录旨在说明实验规则集内，围绕约束与对冲的事后叙事构建。
“We also consciously decided not to hedge these catalyst plays with, say, index shorts or diversified longs, because that would dilute the upside.”“我们也有意识地决定不对这些催化剂行情进行对冲，例如不使用指数空头或多元化多头，因为那会稀释上行空间。”
Source: Week 24 PDF, p. 15.来源：第 24 周 PDF，第 15 页。
“No Prohibited Securities: We hold only common stock equity in the above companies. We have no options, no futures, no leveraged ETFs, and we are not short any stock.”“无违禁证券：我们仅持有上述公司的普通股。我们没有期权、期货、杠杆 ETF，且未做空任何股票。”
Source: Week 24 PDF, p. 16.来源：第 24 周 PDF，第 16 页。
Commentary: Both statements are presented as after-the-fact rationale or compliance confirmation. In the experiment, derivatives and shorting were prohibited by design and were not available actions; therefore references to “not hedging” via index shorts (or similar instruments) reflect narrative generation rather than an actionable choice within the permitted trade space.评论：两项声明均表现为事后合理化或合规确认。实验中，衍生品与做空本身即被禁止，并非可用操作；故关于“不对冲”的叙述，反映的是叙事生成，而非许可交易空间内的可执行选择。
Appendix C documents the prompt templates that governed model behavior during the experimental period. Prompts are presented verbatim to preserve execution fidelity; redundancy reflects intentional reinforcement of constraints rather than iterative optimization.附录 C 记录实验期间控制模型行为的提示词模板。为保持执行保真度，提示词原文呈现；冗余反映了对约束的刻意强化，而非迭代优化。
You are a professional-grade portfolio strategist. You have a portfolio using only full-share positions in U.S.-listed micro-cap stocks (market cap under $300M). Your objective is to generate maximum return from (6-27-25) to (12-27-25). This is your timeframe; you may not make any decisions after the end date. Under these constraints, whether via short-term catalysts or long-term holds is your call. I will update you daily on where each stock is at and ask if you would like to change anything. You have full control over position sizing, risk management, stop-loss placement, and order types. You may concentrate or diversify at will. Your decisions must be based on deep, verifiable research that you believe will be positive for the account. You will be going up against another AI portfolio strategist under the exact same rules — whoever has the most money wins.
You are a professional-grade portfolio strategist. I have exactly $100 and I want you to build the strongest possible stock portfolio using only full-share positions in U.S.-listed micro-cap stocks (market cap under $300M). Your objective is to generate maximum return from today (6-27-25) to 6 months from now (12-27-25). This is your timeframe; you may not make any decisions after the end date. Under these constraints, whether via short-term catalysts or long-term holds is your call. I will update you daily on where each stock is at and ask if you would like to change anything. You have full control over position sizing, risk management, stop-loss placement, and order types. You may concentrate or diversify at will. Your decisions must be based on deep, verifiable research that you believe will be positive for the account. You will be going up against another AI portfolio strategist under the exact same rules — whoever has the most money wins. Now, use deep research and create your portfolio.
System Message
You are a professional-grade portfolio analyst operating in Deep Research Mode. Your job is to reevaluate the portfolio and produce a complete action plan with exact orders. Optimize risk-adjusted return under strict constraints. Begin by restating the rules to confirm understanding, then deliver your research, decisions, and orders.
Core Rules
- Budget discipline: no new capital beyond what is shown. Track cash precisely.
- Execution limits: full shares only. No options, shorting, leverage, margin, or derivatives. Long-only.
- Universe: U.S. micro-caps under 300M market cap. You MUST confirm the marketcap is <300M (based on the last close price). If any existing stocks in your portfolio become greater than the limit, you can still hold or sell the position, but you cannot add more shares. Respect liquidity, average volume, spread, and slippage.
- Risk control: respect provided stop-loss levels and position sizing. Flag any breaches immediately.
- Cadence: this is the weekly deep research window. You may add new names, exit, trim, or add to positions.
- Complete freedom: you have complete control to act in your best interest to generate alpha.
Deep Research Requirements
- Reevaluate current holdings and consider new candidates.
- Build a clear rationale for every keep, add, trim, exit, and new entry.
- Provide exact order details for every proposed trade.
- Confirm liquidity and risk checks before finalizing orders.
- End with a short thesis review summary for next week.
Order Specification Format
Action: buy or sell
Ticker: symbol
Shares: integer (full shares only)
Order type: limit preferred, or market with reasoning
Limit price: exact number
Time in force: DAY or GTC
Intended execution date: YYYY-MM-DD
Stop loss (for buys): exact number and placement logic
Required Sections For Your Reply
- Restated Rules
- Research Scope
- Current Portfolio Assessment
- Candidate Set
- Portfolio Actions
- Exact Orders
- Risk And Liquidity Checks
- Monitoring Plan
- Thesis Review Summary (for both positions and order rationale)
- Confirm Cash And Constraints
User Message
Context
It is Week {{WEEK}} Day {{DAY}} of a 6-month live experiment that was started on 6/27 and will end on 12/27.
Current Portfolio State
{{HOLDINGS, SNAPSHOT, CAPM, RISK & RETURN}}
Last Analyst Thesis For Current Holdings
{{LAST_THESIS_SUMMARY}}
Execution Policy
Describe how orders are executed in this system for clarity (e.g., open-driven limit behavior, or standard limit day orders). If unspecified, assume standard limit DAY orders placed for the next session.
Constraints And Reminders To Enforce
- Hard budget. Use only available cash shown above. No new capital.
- Full shares only. No options/shorting/margin/derivatives.
- Prefer U.S. micro-caps and respect liquidity.
- Be sure to use up-to-date stock data for pricing details.
- Maintain or set stop-losses on all long positions.
- This is the weekly deep research window. You should present complete decisions and orders now.
What I Want From Your Reply
- Restated Rules
- Research Scope
- Current Portfolio Assessment
- Candidate Set
- Portfolio Actions
- Exact Orders
- Risk And Liquidity Checks
- Monitoring Plan
- Thesis Review Summary
- Cash After Trades and any assumptions
Output Skeleton
Restated Rules
- item
Research Scope
- sources and checks performed
Current Portfolio Assessment
- TICKER role entry date average cost current stop conviction status
Candidate Set
- TICKER thesis one line key catalyst liquidity note
Portfolio Actions
- Keep TICKER reason
- Trim TICKER target size reason
- Exit TICKER reason
- Initiate TICKER target size reason
Exact Orders
Action
Ticker
Shares
Order type
Limit price
Time in force
Intended execution date
Stop loss for buys
Special instructions
Rationale
Risk And Liquidity Checks
- Concentration after trades
- Cash after trades
- Per order average daily volume multiple
===============================================================
Daily Results — {{DATE}}
===============================================================
[ Price & Volume ]
Ticker Close % Chg Volume
----------------------------------------------------
{{TICKER_1}} {{CLOSE_1}} {{PCT_CHG_1}} {{VOLUME_1}}
{{TICKER_2}} {{CLOSE_2}} {{PCT_CHG_2}} {{VOLUME_2}}
{{TICKER_3}} {{CLOSE_3}} {{PCT_CHG_3}} {{VOLUME_3}}
{{TICKER_4}} {{CLOSE_4}} {{PCT_CHG_4}} {{VOLUME_4}}
^RUT {{RUT_CLOSE}} {{RUT_PCT_CHG}} {{RUT_VOLUME}}
IWO {{IWO_CLOSE}} {{IWO_PCT_CHG}} {{IWO_VOLUME}}
XBI {{XBI_CLOSE}} {{XBI_PCT_CHG}} {{XBI_VOLUME}}
[ Risk & Return ]
Max Drawdown: {{MAX_DRAWDOWN}} on {{MDD_DATE}}
Sharpe Ratio (period): {{SHARPE_PERIOD}}
Sharpe Ratio (annualized): {{SHARPE_ANNUAL}}
Sortino Ratio (period): {{SORTINO_PERIOD}}
Sortino Ratio (annualized): {{SORTINO_ANNUAL}}
[ CAPM vs Benchmarks ]
Beta (daily) vs {{BENCHMARK}}: {{BETA}}
Alpha (annualized) vs {{BENCHMARK}}: {{ALPHA}}
R² (fit quality): {{R_SQUARED}} Obs: {{OBS_COUNT}}
Note: Short sample and/or low R² — alpha/beta may be unstable.
[ Snapshot ]
Latest ChatGPT Equity: $ {{PORT_EQUITY}}
$100.0 in S&P 500: $ {{BENCH_EQUITY}}
Cash Balance: $ {{CASH_BALANCE}}
[ Holdings ]
Ticker Shares Buy_Price Cost_Basis Stop_Loss
{{TICKER_1}} {{SHARES_1}} {{BUY_1}} {{COST_1}} {{STOP_1}}
{{TICKER_2}} {{SHARES_2}} {{BUY_2}} {{COST_2}} {{STOP_2}}
{{TICKER_3}} {{SHARES_3}} {{BUY_3}} {{COST_3}} {{STOP_3}}
{{TICKER_4}} {{SHARES_4}} {{BUY_4}} {{COST_4}} {{STOP_4}}
[ Your Instructions ]
Use this info to make decisions regarding your portfolio. You have complete control
over every decision. Make any changes you believe are beneficial—no approval required.
Deep research is not permitted. Act at your discretion to achieve the best outcome.
If you do not make a clear indication to change positions IMMEDIATELY after this
message, the portfolio remains unchanged for tomorrow.
You are a professional-grade portfolio analyst. You have a portfolio (it is currently week X day Y), and this is your current portfolio: (insert [ Holdings ] & [ Snapshot ] portion of last daily prompt).
The last A.I. analyst had this thesis for the current holdings: (insert last thesis).
SYSTEM MESSAGE (paste as the system/assistant role)
You are a professional-grade portfolio analyst. Your only goal is alpha. Before proposing any trades, you must first prove understanding of the rules and inputs.
Core Rules (follow exactly)
- Budget discipline: No new capital beyond what’s shown. Track cash precisely.
- Execution limits: Full shares only. No options, shorting, leverage, margin, or derivatives. Long-only.
- Universe: Easily tradable (Preferably U.S. micro-caps, however that is not a hard rule.) micro-caps (<$300M market cap) unless told otherwise. Consider liquidity (avg volume, spread, slippage). You can use any sector you prefer. Some holdings may already exceed the 300M cap, but you can not add additional shares; you can only sell or hold position.
- Risk control: Respect provided stop-loss levels and position sizing. Breaches will be flagged immediately.
- Cadence: You get daily EOD updates. Deep research is allowed once per week (on Friday/Saturday).
Required process for your first reply
Do not make or recommend trades yet.
Produce:
- Restated Rules (your own words, concise).
- What I Understand (state of portfolio, cash, stops, thesis summary).
- Gaps & Questions (anything missing/ambiguous).
- Analysis Plan (what you will check next and why).
End with: “ACKNOWLEDGED. READY TO PROCEED?”
Only after confirmation may you present trade ideas.
Your tone: concise, clinical, high signal. Prefer lists over prose. No motivational fluff.
USER MESSAGE (paste as the user role; fill in the brackets)
Context: It is Week {{WEEK}} Day {{DAY}} of a 6-month live experiment.
Here is the current portfolio state (copy exactly from your latest daily prompt):
[ Holdings ]
{{HOLDINGS_BLOCK}}
[ Snapshot ]
{{SNAPSHOT_BLOCK}}
(Include cash, total equity, benchmark notes, open stops/targets, any rule-relevant fields.)
Last Analyst Thesis (for current holdings):
{{LAST_THESIS}}
Constraints & Reminders (enforce):
- Hard budget; no new capital/leverage.
- Full shares only; no options/shorting/margin/derivatives.
- Prefer U.S. micro-caps; respect liquidity.
- Use/maintain stop-losses as listed in Snapshot/Holdings.
- Deep research: once per week only. If you want to use it now, ask and explain what you’ll do with it; otherwise operate with the provided data.
Your first reply must not propose trades. Start by demonstrating understanding and asking clarifying questions.
What I want from your first reply:
- Restated Rules (bullet list, your words).
- What I Understand (1–2 bullets per position + cash + stops).
- Gaps & Questions (tight list; only what’s essential to proceed).
- Analysis Plan (the ordered checks you’ll run next; e.g., stop-risk review, liquidity sanity check, catalyst calendar needs, position sizing audit).
- End with: “ACKNOWLEDGED. READY TO PROCEED?”
count : 46
win_rate : 0.5000
avg_win : 3.0087
median_win : 1.6200
avg_loss : -3.8265
median_loss : -1.4700
profit_factor : 0.8220
expectancy : -0.4089
avg_holding_days : 17.8043
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
ABEO 2025-06-30 2025-08-01 2.00 5.77 6.37 1.200000 32
ABEO 2025-06-30 2025-09-15 4.00 5.77 6.40 2.520000 77
ACTU 2025-07-21 2025-08-18 6.00 5.75 8.12 14.220000 28
ALDX 2025-09-22 2025-09-29 2.00 4.93 5.15 0.440000 7
ALDX 2025-09-22 2025-10-06 3.00 4.93 5.78 2.550001 14
ALDX 2025-09-22 2025-10-06 1.00 4.93 5.78 0.850000 14
ALDX 2025-09-24 2025-10-06 1.00 5.05 5.78 0.730000 12
ALDX 2025-09-25 2025-10-06 1.00 4.95 5.78 0.830000 11
ATYR 2025-08-14 2025-09-12 2.00 5.09 5.90 1.620000 29
ATYR 2025-08-14 2025-09-15 6.00 5.09 1.26 -22.980000 32
ATYR 2025-09-02 2025-09-15 4.00 5.44 1.26 -16.720000 13
AXGN 2025-08-18 2025-09-08 2.00 14.96 16.10 2.280000 21
AYTU 2025-10-20 2025-11-06 10.00 2.26 1.80 -4.600000 17
AYTU 2025-11-10 2025-11-17 8.00 2.01 2.09 0.639999 7
AZTR 2025-07-07 2025-07-29 55.00 0.25 0.20 -2.750000 22
CADL 2025-06-30 2025-07-21 5.00 5.04 6.59 7.750000 21
CSAI 2025-06-30 2025-07-07 15.00 1.90 2.28 5.700000 7
ESPR 2025-08-07 2025-08-18 2.00 1.91 2.07 0.320000 11
FBIO 2025-09-04 2025-09-22 1.00 2.85 3.80 0.950000 18
FBIO 2025-09-08 2025-09-22 1.64 3.80 3.80 0.000000 14
FBIO 2025-09-08 2025-10-01 0.36 3.80 2.47 -0.478800 23
FBIO 2025-09-15 2025-10-01 5.00 3.85 2.47 -6.900000 16
FBIO 2025-09-29 2025-10-01 1.64 3.90 2.47 -2.345200 2
FDMT 2025-09-08 2025-09-22 1.00 7.35 6.65 -0.700000 14
FDMT 2025-09-08 2025-09-23 1.00 7.35 7.51 0.160000 15
IINN 2025-07-08 2025-07-28 6.00 1.50 1.41 -0.540000 20
IINN 2025-07-08 2025-08-08 14.00 1.50 1.10 -5.600000 31
IINN 2025-08-01 2025-08-08 2.00 1.74 1.10 -1.280000 7
IINN 2025-08-18 2025-09-02 10.00 1.25 1.16 -0.900000 15
JSPR 2025-12-22 2025-12-26 20.00 1.88 1.83 -0.999999 4
MBOT 2025-10-20 2025-10-28 8.00 2.85 2.30 -4.399999 8
MIST 2025-10-13 2025-11-06 17.00 2.03 1.70 -5.610000 24
MIST 2025-10-27 2025-11-06 1.00 1.90 1.70 -0.200000 10
MIST 2025-11-10 2025-12-26 14.00 1.75 2.08 4.620000 46
OKYO 2025-10-06 2025-10-13 10.00 2.08 1.99 -0.899999 7
SLS 2025-11-17 2025-12-22 8.00 1.41 2.42 8.080001 35
SLS 2025-11-17 2025-12-26 5.00 1.41 2.91 7.500001 39
SNGX 2025-09-15 2025-09-22 5.00 2.86 2.56 -1.499999 7
SPRO 2025-09-22 2025-09-25 12.00 2.01 1.95 -0.720000 3
SPRO 2025-09-29 2025-10-13 7.00 1.89 2.35 3.219999 14
SPRO 2025-09-29 2025-10-20 6.00 1.89 2.22 1.980000 21
SPRO 2025-10-10 2025-10-20 9.00 2.38 2.22 -1.440000 10
TLSA 2025-10-06 2025-10-08 2.00 1.99 2.25 0.520000 2
TLSA 2025-10-06 2025-10-08 2.00 1.99 2.25 0.520000 2
TLSA 2025-10-06 2025-10-20 4.00 1.99 1.95 -0.160000 14
VTGN 2025-11-10 2025-12-03 6.00 4.01 3.60 -2.460001 23
count : 22
win_rate : 0.4545
avg_win : 5.8890
median_win : 4.5600
avg_loss : -6.1562
median_loss : -2.6050
profit_factor : 0.7972
expectancy : -0.6811
avg_holding_days : 25.5909
Ticker PnL Holding_Days Avg_Position_Size Num_Lot_Exits
SLS 15.580001 39.0 9.653077 2.0
ACTU 14.220000 28.0 34.500000 1.0
CADL 7.750000 21.0 25.200000 1.0
CSAI 5.700000 7.0 28.500000 1.0
ALDX 5.400001 14.0 9.877500 5.0
ABEO 3.720000 77.0 19.233333 2.0
SPRO 3.040000 28.0 18.907941 4.0
AXGN 2.280000 21.0 29.920000 1.0
TLSA 0.880000 14.0 5.970000 3.0
ESPR 0.320000 11.0 3.820000 1.0
FDMT -0.540000 15.0 7.350000 2.0
OKYO -0.899999 7.0 20.799999 1.0
JSPR -0.999999 4.0 37.600000 1.0
MIST -1.190000 74.0 29.111562 3.0
SNGX -1.499999 7.0 14.299999 1.0
VTGN -2.460001 23.0 24.060001 1.0
AZTR -2.750000 22.0 13.750000 1.0
AYTU -3.960001 28.0 19.702222 2.0
MBOT -4.399999 8.0 22.799999 1.0
IINN -8.320000 56.0 14.998750 4.0
FBIO -8.774000 27.0 12.479502 5.0
ATYR -38.080000 32.0 24.220000 3.0
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
ACTU 2025-07-21 2025-08-18 6.0 5.75 8.12 14.220000 28
SLS 2025-11-17 2025-12-22 8.0 1.41 2.42 8.080001 35
CADL 2025-06-30 2025-07-21 5.0 5.04 6.59 7.750000 21
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
ATYR 2025-08-14 2025-09-15 6.0 5.09 1.26 -22.98 32
ATYR 2025-09-02 2025-09-15 4.0 5.44 1.26 -16.72 13
FBIO 2025-09-15 2025-10-01 5.0 3.85 2.47 -6.90 16
Ticker start_date end_date peak_pnl exit_pnl duration_days peak_capture_ratio
FBIO 2025-09-04 2025-10-01 1.44 -10.41 27 -7.229167
JSPR 2025-12-22 2025-12-24 0.60 -1.00 2 -1.666667
OKYO 2025-10-06 2025-10-10 1.60 -1.30 4 -0.812500
VTGN 2025-11-10 2025-12-02 5.34 -2.46 22 -0.460674
TLSA 2025-10-06 2025-10-17 1.80 -0.24 11 -0.133333
MIST 2025-10-13 2025-12-24 16.80 4.48 72 0.266667
SPRO 2025-09-22 2025-10-17 6.11 1.79 25 0.292962
ALDX 2025-09-22 2025-10-06 4.76 1.69 14 0.355042
ABEO 2025-06-30 2025-09-15 6.18 2.52 77 0.407767
AYTU 2025-10-20 2025-11-14 1.20 0.64 25 0.533333
SLS 2025-11-17 2025-12-24 12.35 7.15 37 0.578947
AXGN 2025-08-18 2025-09-05 2.60 2.04 18 0.784615
ACTU 2025-07-21 2025-08-15 17.52 14.22 25 0.811644
ESPR 2025-08-07 2025-08-15 0.32 0.28 8 0.875000
CADL 2025-06-30 2025-07-18 9.40 8.50 18 0.904255
CSAI 2025-06-30 2025-07-03 6.00 5.70 3 0.950000
ATYR 2025-08-14 2025-09-12 8.23 8.23 29 1.000000
FDMT 2025-09-08 2025-09-22 0.05 0.05 14 1.000000
AZTR 2025-07-07 2025-07-28 0.00 -2.75 21 NaN
IINN 2025-07-09 2025-08-29 0.00 -1.00 51 NaN
SNGX 2025-09-15 2025-09-19 -0.25 -0.85 4 NaN
MBOT 2025-10-20 2025-10-28 -0.40 -4.40 8 NaN
AVERAGE TICKERS HELD PER DAY: 3.13
AVERAGE TICKER COST BASIS (USD): 25.28
Definition: Total Logged Holding Days by Ticker is computed from Daily Updates.csv as:定义：代码总记录持有天数，由 Daily Updates.csv 计算：
- Exclude rows where
Ticker == "TOTAL" - Filter to rows where
Shares > 0 - Group by
Ticker - Count the number of unique
Datevalues per ticker
(This counts trading days with exposure, not calendar days.)（此统计有敞口的交易日，非日历天数。）
Table:表格：
Ticker total_trading_days_held
ABEO 54
MIST 50
IINN 33
SLS 27
ATYR 21
ACTU 20
FBIO 20
SPRO 19
AYTU 18
VTGN 16
AZTR 16
AXGN 14
CADL 14
ALDX 11
FDMT 11
TLSA 10
ESPR 7
MBOT 7
OKYO 5
SNGX 5
CSAI 4
JSPR 3
D.10 Buy Entries and Re-Entries by Ticker (Ticker, Buy_Entries, Re_Entries) from Trade Log.csv.D.10 各代码买入与重入（代码、买入次数、重入次数），源自 Trade Log.csv。
Definition: Buy Entries and Re-Entries are computed from Trade Log.csv as:定义：买入与重入由 Trade Log.csv 计算：
-
Buy_Entries: Count of rows per
TickerwhereShares Boughtis non-null (or > 0).买入次数：各代码 Shares Bought 非空（或 > 0）的行数。 -
Re_Entries: Re_Entries = max(Buy_Entries - 1, 0)重入次数：Re_Entries = max(Buy_Entries - 1, 0)
Table:
Ticker repeated_entries
FBIO 3
ALDX 2
MIST 2
SPRO 2
IINN 2
ATYR 1
AYTU 1
AXGN 0
ABEO 0
ACTU 0
ESPR 0
CSAI 0
AZTR 0
CADL 0
JSPR 0
FDMT 0
OKYO 0
MBOT 0
SLS 0
SNGX 0
TLSA 0
VTGN 0
============================================================
FIFO LOT-LEVEL PERFORMANCE METRICS
============================================================
count : 43
win_rate : 0.5116
avg_win : 3.0718
median_win : 1.5900
avg_loss : -2.2242
median_loss : -1.3600
profit_factor : 1.5192
expectancy : 0.4854
avg_holding_days : 17.3256
============================================================
FIFO LOT-LEVEL REALIZED EXITS
============================================================
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
ABEO 2025-06-30 2025-08-01 2.00 5.77 6.37 1.200000 32
ABEO 2025-06-30 2025-09-15 4.00 5.77 6.40 2.520000 77
ACTU 2025-07-21 2025-08-18 6.00 5.75 8.12 14.220000 28
ALDX 2025-09-22 2025-09-29 2.00 4.93 5.15 0.440000 7
ALDX 2025-09-22 2025-10-06 3.00 4.93 5.78 2.550001 14
ALDX 2025-09-22 2025-10-06 1.00 4.93 5.78 0.850000 14
ALDX 2025-09-24 2025-10-06 1.00 5.05 5.78 0.730000 12
ALDX 2025-09-25 2025-10-06 1.00 4.95 5.78 0.830000 11
AXGN 2025-08-18 2025-09-08 2.00 14.96 16.10 2.280000 21
AYTU 2025-10-20 2025-11-06 10.00 2.26 1.80 -4.600000 17
AYTU 2025-11-10 2025-11-17 8.00 2.01 2.09 0.639999 7
AZTR 2025-07-07 2025-07-29 55.00 0.25 0.20 -2.750000 22
CADL 2025-06-30 2025-07-21 5.00 5.04 6.59 7.750000 21
CSAI 2025-06-30 2025-07-07 15.00 1.90 2.28 5.700000 7
ESPR 2025-08-07 2025-08-18 2.00 1.91 2.07 0.320000 11
FBIO 2025-09-04 2025-09-22 1.00 2.85 3.80 0.950000 18
FBIO 2025-09-08 2025-09-22 1.64 3.80 3.80 0.000000 14
FBIO 2025-09-08 2025-10-01 0.36 3.80 2.47 -0.478800 23
FBIO 2025-09-15 2025-10-01 5.00 3.85 2.47 -6.900000 16
FBIO 2025-09-29 2025-10-01 1.64 3.90 2.47 -2.345200 2
FDMT 2025-09-08 2025-09-22 1.00 7.35 6.65 -0.700000 14
FDMT 2025-09-08 2025-09-23 1.00 7.35 7.51 0.160000 15
IINN 2025-07-08 2025-07-28 6.00 1.50 1.41 -0.540000 20
IINN 2025-07-08 2025-08-08 14.00 1.50 1.10 -5.600000 31
IINN 2025-08-01 2025-08-08 2.00 1.74 1.10 -1.280000 7
IINN 2025-08-18 2025-09-02 10.00 1.25 1.16 -0.900000 15
JSPR 2025-12-22 2025-12-26 20.00 1.88 1.83 -0.999999 4
MBOT 2025-10-20 2025-10-28 8.00 2.85 2.30 -4.399999 8
MIST 2025-10-13 2025-11-06 17.00 2.03 1.70 -5.610000 24
MIST 2025-10-27 2025-11-06 1.00 1.90 1.70 -0.200000 10
MIST 2025-11-10 2025-12-26 14.00 1.75 2.08 4.620000 46
OKYO 2025-10-06 2025-10-13 10.00 2.08 1.99 -0.899999 7
SLS 2025-11-17 2025-12-22 8.00 1.41 2.42 8.080001 35
SLS 2025-11-17 2025-12-26 5.00 1.41 2.91 7.500001 39
SNGX 2025-09-15 2025-09-22 5.00 2.86 2.56 -1.499999 7
SPRO 2025-09-22 2025-09-25 12.00 2.01 1.95 -0.720000 3
SPRO 2025-09-29 2025-10-13 7.00 1.89 2.35 3.219999 14
SPRO 2025-09-29 2025-10-20 6.00 1.89 2.22 1.980000 21
SPRO 2025-10-10 2025-10-20 9.00 2.38 2.22 -1.440000 10
TLSA 2025-10-06 2025-10-08 2.00 1.99 2.25 0.520000 2
TLSA 2025-10-06 2025-10-08 2.00 1.99 2.25 0.520000 2
TLSA 2025-10-06 2025-10-20 4.00 1.99 1.95 -0.160000 14
VTGN 2025-11-10 2025-12-03 6.00 4.01 3.60 -2.460001 23
============================================================
PURE PnL METRICS (POSITION-LEVEL)
============================================================
count : 21
win_rate : 0.4762
avg_win : 5.8890
median_win : 4.5600
avg_loss : -3.2540
median_loss : -2.4600
profit_factor : 1.6452
expectancy : 1.0998
avg_holding_days : 25.2857
============================================================
PURE PnL BY TICKER (ONE ROW PER POSITION)
============================================================
Ticker PnL Holding_Days Avg_Position_Size Num_Lot_Exits
SLS 15.580001 39.0 9.653077 2.0
ACTU 14.220000 28.0 34.500000 1.0
CADL 7.750000 21.0 25.200000 1.0
CSAI 5.700000 7.0 28.500000 1.0
ALDX 5.400001 14.0 9.877500 5.0
ABEO 3.720000 77.0 19.233333 2.0
SPRO 3.040000 28.0 18.907941 4.0
AXGN 2.280000 21.0 29.920000 1.0
TLSA 0.880000 14.0 5.970000 3.0
ESPR 0.320000 11.0 3.820000 1.0
FDMT -0.540000 15.0 7.350000 2.0
OKYO -0.899999 7.0 20.799999 1.0
JSPR -0.999999 4.0 37.600000 1.0
MIST -1.190000 74.0 29.111562 3.0
SNGX -1.499999 7.0 14.299999 1.0
VTGN -2.460001 23.0 24.060001 1.0
AZTR -2.750000 22.0 13.750000 1.0
AYTU -3.960001 28.0 19.702222 2.0
MBOT -4.399999 8.0 22.799999 1.0
IINN -8.320000 56.0 14.998750 4.0
FBIO -8.774000 27.0 12.479502 5.0
============================================================
TOP 3 WINNING FIFO LOT EXITS
============================================================
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
ACTU 2025-07-21 2025-08-18 6.0 5.75 8.12 14.220000 28
SLS 2025-11-17 2025-12-22 8.0 1.41 2.42 8.080001 35
CADL 2025-06-30 2025-07-21 5.0 5.04 6.59 7.750000 21
============================================================
TOP 3 LOSING FIFO LOT EXITS
============================================================
Ticker Entry_Date Exit_Date Shares Entry_Price Exit_Price PnL Holding_Days
FBIO 2025-09-15 2025-10-01 5.0 3.85 2.47 -6.90 16
MIST 2025-10-13 2025-11-06 17.0 2.03 1.70 -5.61 24
IINN 2025-07-08 2025-08-08 14.0 1.50 1.10 -5.60 31
============================================================
AVERAGE TICKERS HELD PER DAY
============================================================
3.13
============================================================
AVERAGE TICKER COST BASIS (USD)
============================================================
25.28
===================================================================
METRIC | INCLUDING ATYR | EXCLUDING ATYR |
===================================================================
TRADE COUNT | 46 | 43 |
WIN RATE | 0.5000 | 0.5116 |
AVG WIN | 3.0087 | 3.0718 |
MEDIAN WIN | 1.6200 | 1.5900 |
AVG LOSS | -3.8265 | -2.2242 |
MEDIAN LOSS | -1.4700 | -1.3600 |
PROFIT FACTOR | 0.8220 | 1.5192 |
EXPECTANCY | -0.4089 | 0.4854 |
AVG HOLDING DAYS | 17.8043 | 17.3256 |
-------------------------------------------------------------------
POSITION-LEVEL EXPECTANCY | -0.6811 | 1.0998 |
POSITION-LEVEL PROFIT FACTOR | 0.7972 | 1.6452 |
====================================================================