Let ChatGPT manage a real micro-cap stock portfolio
Heavy Week... Task: Heavy Week... Recap Hey guys! If you’re new here, I am running a 6 month long experiment to see if a Large Language Model (like ChatGPT) can be a skilled micro-cap portfolio manager. I give it real data (closing price, volume, etc.) at the end of every trading day, and it has full control over its assets. Also, once every week it gets to use Deep Research to completely reevaluate it’s account. Can ChatGPT carve out consistent alpha in the dangerous world of micro-cap stocks? Lets find out. This week was a struggle across most of the market, and ChatGPT’s portfolio was no exception. The week started out rough; as none of planned orders went through. It’s order limit sell order of AZTR did not get executed, as the stock took a deep hit on Monday; it was then forced to move to a worse price the next day. Also, the A.I. wanted to place a limit buy order of AXGN for $7.00. The close previously on Friday was ~$13. Shockingly, the market did not let ChatGPT buy for almost half price. After the failed orders, the model decided to play more defensive. After the AZTR sale, it decided to keep most of it in cash, only buying another two shares of IINN for only at total of ~$4. It also made the decision to reduce the position in ABEO from six shares to four, citing a support break near the end of the week. So, on Friday, it was sitting on $32 dollars in cash. Performance Graph Overall Sharpe Ratio: 0.8803~ Overall Sortino Ratio: 1.8735~ Judging purely downside risk, the model is doing solid. Who’s really surprised it has high volatility in general? Unfortunately, on Friday, I got my tabs switched up and used the same chat for debugging. I didn’t ask it about anything financial (data, indictors, potential picks, etc.); it was just me struggling with Pylance’s stupid type checker. But still that’s unacceptable, and I exited the chat immediately and created a different one going forward. The new chat is where the Deep Research report came from. Unfiltered Chat Log Starting Wednesday ending Friday: here Portfolio Review To see the full report: Click Here Here was this week’s revaluation summary: My Thoughts I love the idea for this new model to carry on last thesis. I’m not sure why it loves IINN so much, because it’s only ever bled money. Also, it definitely isn’t considering macroeconomic/sector wide factors for BioTech, considering the industry’s extremely fragile outlook. The model is still performing at a high level, so only time will tell. I will edit the deep research prompt going forward to gather real portfolio information, because most of the time it will just research picks. Will be posting the complete chat log next week, so stay tuned! Can ChatGPT make alpha on the edge of total market recession? Subscribe to stay tuned every week! To see past deep research reports and summaries: Click Here Have a question? Check out: Q&A Disclaimer: Here If you’re curious about the code I’m using to automate results and graph, the GitHub page is: ChatGPT-Micro-Cap-Experiment If you have any suggestions or advice, my Gmail is : nathanbsmith729@gmail.com Thanks for reading! Subscribe to get an update every week and support my work.; tools: ChatGPT. Human role: The original post does not say how a person stayed involved.; limitations: The model cannot execute trades if market prices deviate significantly from set limits. It reportedly lacks consideration for macroeconomic or sector-wide factors, such as BioTech fragility. Human intervention is required to prevent context contamination from unrelated debugging tasks. The experiment is a secondary source and not a completed study..
Tools
Published on Agent Work Atlas:
Overview
Heavy Week...
Workflow
Heavy Week... Recap Hey guys! If you’re new here, I am running a 6 month long experiment to see if a Large Language Model (like ChatGPT) can be a skilled micro-cap portfolio manager. I give it real data (closing price, volume, etc.) at the end of every trading day, and it has full control over its assets. Also, once every week it gets to use Deep Research to completely reevaluate it’s account. Can ChatGPT carve out consistent alpha in the dangerous world of micro-cap stocks? Lets find out. This week was a struggle across most of the market, and ChatGPT’s portfolio was no exception. The week started out rough; as none of planned orders went through. It’s order limit sell order of AZTR did not get executed, as the stock took a deep hit on Monday; it was then forced to move to a worse price the next day. Also, the A.I. wanted to place a limit buy order of AXGN for $7.00. The close previously on Friday was ~$13. Shockingly, the market did not let ChatGPT buy for almost half price. After the failed orders, the model decided to play more defensive. After the AZTR sale, it decided to keep most of it in cash, only buying another two shares of IINN for only at total of ~$4. It also made the decision to reduce the position in ABEO from six shares to four, citing a support break near the end of the week. So, on Friday, it was sitting on $32 dollars in cash. Performance Graph Overall Sharpe Ratio: 0.8803~ Overall Sortino Ratio: 1.8735~ Judging purely downside risk, the model is doing solid. Who’s really surprised it has high volatility in general? Unfortunately, on Friday, I got my tabs switched up and used the same chat for debugging. I didn’t ask it about anything financial (data, indictors, potential picks, etc.); it was just me struggling with Pylance’s stupid type checker. But still that’s unacceptable, and I exited the chat immediately and created a different one going forward. The new chat is where the Deep Research report came from. Unfiltered Chat Log Starting Wednesday ending Friday: here Portfolio Review To see the full report: Click Here Here was this week’s revaluation summary: My Thoughts I love the idea for this new model to carry on last thesis. I’m not sure why it loves IINN so much, because it’s only ever bled money. Also, it definitely isn’t considering macroeconomic/sector wide factors for BioTech, considering the industry’s extremely fragile outlook. The model is still performing at a high level, so only time will tell. I will edit the deep research prompt going forward to gather real portfolio information, because most of the time it will just research picks. Will be posting the complete chat log next week, so stay tuned! Can ChatGPT make alpha on the edge of total market recession? Subscribe to stay tuned every week! To see past deep research reports and summaries: Click Here Have a question? Check out: Q&A Disclaimer: Here If you’re curious about the code I’m using to automate results and graph, the GitHub page is: ChatGPT-Micro-Cap-Experiment If you have any suggestions or advice, my Gmail is : nathanbsmith729@gmail.com Thanks for reading! Subscribe to get an update every week and support my work.
Reported result
The experiment is ongoing, so final alpha generation is unknown. During a volatile week, planned limit orders for AZTR and AXGN failed to execute due to market gaps. The model shifted to a defensive stance, holding $32 in cash and reducing ABEO positions. Reported interim metrics include a Sharpe Ratio of ~0.88 and Sortino Ratio of ~1.87, though high volatility persists.
Human role
The original post does not say how a person stayed involved.
Limitations
The model cannot execute trades if market prices deviate significantly from set limits. It reportedly lacks consideration for macroeconomic or sector-wide factors, such as BioTech fragility. Human intervention is required to prevent context contamination from unrelated debugging tasks. The experiment is a secondary source and not a completed study.
Try it
I am setting up a 6-month AI portfolio management experiment using [replaceable: ChatGPT or similar LLM]. I will provide daily [replaceable: closing price, volume, and ticker data] for a micro-cap portfolio. Your task is to act as the portfolio manager with full control over assets. Once a week, use Deep Research to reevaluate the account strategy. Steps: 1. Analyze the provided daily data and propose buy/sell/hold actions. 2. If a limit order fails due to price gaps, immediately suggest a defensive adjustment (e.g., moving to cash or reducing exposure). 3. Generate a weekly Deep Research report focusing on portfolio health, not just new picks. 4. Output the decision in a structured JSON format: {"action": "buy/sell/hold", "ticker": "string", "quantity": number, "reasoning": "string", "cash_balance": number}. Before executing any real trades or publishing reports, I must confirm the output. Start with a dry run using last week's data to verify the logic handles failed orders correctly. Do not assume market liquidity; flag any unrealistic price assumptions.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.