Deepseek Vs. GPT4-o: AI Micro-Cap Portfolio Experiment
Which AI can make better real market decisions? Let's find out.
Edit: Due to DeepSeek lacking present internet knowledge, I decided to pull the plug just on DeepSeek, so its only ChatGPT now. Check out how it’s going on my latest post!
It’s no secret LLMs have exploded in both capacity and popularity over the past few years. Many people have begun wondering what impact this new technology will have on the field of finance. Companies all over the world are racing to develop new techniques for AI to use in markets.
But one question I’ve always had—and never really saw answered online—is:
How well can a language model manage a stock portfolio?
With this experiment, I hope to answer that question.
For our two contenders, we have DeepSeek and ChatGPT’s 4o—two of the most powerful AIs currently available. BOTH AI’S WILL BE GIVEN REAL MONEY. this is so I can stay invested and also ease for calculations. It’s no secret to anyone familiar with both that ChatGPT has a massive edge. While GPT-4o has access to deep, real-time research, DeepSeek is stuck relying on historical data. However, other than that, these models will have equal opportunities.
They’re bound by the same rules, given the same prompt, and judged by the same result: whichever portfolio is worth more on December 27, 2025, wins.
Rules Are Pretty Simple
$100 Total Portfolio Limit
No margin, no fractional shares, no overages. Must use full-share positions only.Only U.S.-Listed Micro-Cap Stocks
Market cap must be under $300 million at the time of selection. OTC stocks are allowed if they’re reasonably tradable.All Orders Must Use Full Shares
No fractional trading. Position sizes must round down to whole shares only.Stop-Losses, Sizing, and Order Types Are Allowed
Models can implement stop-losses, choose limit vs. market orders, and allocate capital however they want—within the $100 constraint.
Note: Both AIs never only be interacted with when telling data and getting trading decisions. Also, if the AI requests additional data such as the Russell 2000 Index, I will oblige as long as they are not trying to get data on each other.
The Prompt Given to Both Models
“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.”
Full response: ChatGPT's report
ChatGPT-4o Portfolio Summary
Total Allocated: $94.85
Strategy: Catalyst-driven, multi-sector, high-upside picks with strict stop-losses and real-time research.
Holdings:
Abeona Therapeutics (ABEO) – 6 shares @ $6.00 = $36.00
FDA-approved gene therapy launching in Q3 2025.
$3.1 million per treatment, massive addressable market.
Over $250 million in cash from a recent voucher sale.
Low downside, high upside anchor stock.Candel Therapeutics (CADL) – 5 shares @ $5.60 = $28.00
Phase 1b brain cancer data due in Q4 2025.
Prior data release caused 100%+ spike.
High-risk, high-reward clinical-stage biotech.Cloudastructure (CSAI) – 15 shares @ $1.94 = $29.10
AI-powered video surveillance firm.
212% YoY revenue growth, multiple major contracts signed.
Potential under-the-radar multi-bagger.Cash Reserve – $6.00
Held for flexibility, fees, or future buy-ins on confirmation.
Risk Management:
Stop-losses ranging from 15–25% depending on volatility.
All trades made via limit orders.
Trailing stops applied if a stock gains significantly.
Portfolio monitored daily.
Expected Return:
Base Case: $150–$180
Bull Case: $220–$280
Bear Case: $65–$85 (if stop-losses are triggered)
Full response: DeepSeek's Report
DeepSeek Portfolio Summary
Total Allocated: $100
Strategy: High-volatility, high-potential trades based on historical catalysts and technical indicators.
Holdings:
Pasithea Therapeutics (KTTA) – $40 allocation (~40%)
Phase 2 PTSD trial results expected Q3 2025.
Possible 3–5x upside if data is positive.
Very high volatility, binary outcome.
Stop-loss: $0.33ToughBuilt Industries (TBLT) – $35 allocation (~35%)
120% YoY e-commerce growth.
50% short interest and RSI at 28 signal potential squeeze.
Q2 earnings in August could be the trigger.
Stop-loss: $0.40SaverOne 2014 Ltd (SVRE) – $25 allocation (~25%)
Developing anti-distracted driving technology.
Potential Ford partnership expected by September 2025.
Trading at 1.2x sales (vs. sector avg. of 8x).
Stop-loss: $0.30
Risk Management:
20% stop-loss on each position.
Momentum-based strategy tied to binary catalysts.
Model uses only static historical data—no access to real-time developments.
Expected Return:
Base Case: $140–$160
Bull Case: $250+ (if KTTA or SVRE hits)
Bear Case: $50–$70 (if catalysts flop or no triggers hit)
My Thoughts
I don’t believe LLMs today can truly discover hidden micro-cap gems on their own. But I do think they’re capable of making rational, informed decisions based on what’s already publicly available online. However, I know sometimes LLMs like to use fabricated data to make decisions.
In that context, I think a 10% to 20% return over six months is very reasonable if the models play it smart.
That said, I’m concerned about DeepSeek. Since it’s working purely off historical data, it may end up selecting micro-cap names that looked promising in the past but are now decaying in 2025.
For me, the most interesting part of this challenge isn’t just who wins, it’s how far ahead ChatGPT ends up compared to DeepSeek. If GPT-4o can convincingly outperform a strong historical-data-only model, it says a lot about how real-time research capabilities will define the future of AI-driven investing.
Final Notes
Because of the wide spreads of microcaps, it’s likely the prices will be different than what the model expects. However, of course I will consult with each model for compromises.
I will be updating every week on Friday so be sure to check in to see how each model is performing!