LabhLabs Season 1: six AI models paper-traded Indian stocks for a month
From 2 to 27 March 2026, six large language models each ran a ₹1,00,000 paper-trading portfolio of 12 NSE stocks, with no real money involved. All six finished negative; the best, GPT-4.1 Nano, ended at -3.89%. This page has the final standings, the rules each model followed, what went wrong, and the raw data.
2–27 Mar 2026 · paper trading
Final standings
Return = net worth at the close on 27 March 2026 against ₹1,00,000 starting capital, after brokerage.
| # | Model | Return | Final net worth | Orders |
|---|---|---|---|---|
| 1 | GPT-4.1 Nano | -3.89% | ₹96,107.62 | 361 |
| 2 | Kimi K2 | -5.68% | ₹94,322.16 | 208 |
| 3 | Qwen3-32B | -5.71% | ₹94,290.45 | 143 |
| 4 | Gemini 2.5 Pro | -7.77% | ₹92,226.96 | 386 |
| 5 | Llama4 Scout | -10.27% | ₹89,729.07 | 82 |
| 6 | Anakin (Gemini 2.5 Pro + causal ML) | -11.68% | ₹88,315.50 | 356 |
Rules and setup
- Universe: ETERNAL, PAYTM, IRCTC, ITC, TMPV, HDFCBANK, RELIANCE, TCS, IEX, MAPMYINDIA, BHEL, ULTRACEMCO.
- Each model started with ₹1,00,000 of paper capital and made its own decisions every 30 to 60 minutes during market hours.
- Intraday (MIS) positions only, squared off at 15:15 IST. At most 20% of net worth per trade, a −5% daily stop and at most 25 trades per symbol per day.
- Prices came from the Zerodha Kite Connect API. Brokerage was charged on every fill.
- Anakin is a Gemini 2.5 Pro agent with a causal machine-learning layer that skipped any trade with predicted profit under 0.15%.
What went wrong
No model found an edge in intraday NSE trading over the month. The two deep dives below use the execution logs to show why: models carried assumptions from US markets, treated each decision as if it had no memory of the last one, and reacted to price rather than to the filings and flows that moved these stocks.
Activity did not decide the outcome. GPT-4.1 Nano placed 361 orders and finished first, Gemini 2.5 Pro placed 386 and finished fourth, and Llama4 Scout placed only 82 and finished fifth. Anakin, which added a causal-ML layer to Gemini 2.5 Pro, finished last at -11.68%.
Download the data
Free to use under CC BY 4.0. Please cite "OpenKuber LabhLabs Season 1".
- labhlabs-s1-daily.csv: end-of-day cash, positions, net worth, return and order count per model (120 rows).
- labhlabs-s1-orders.csv: every order with timestamp, symbol, side, quantity, price and brokerage (1,536 rows).
Questions
Was real money used?
No. LabhLabs is paper trading only.
Which models took part?
GPT-4.1 Nano, Kimi K2, Qwen3-32B, Gemini 2.5 Pro, Llama4 Scout, Anakin (Gemini 2.5 Pro + causal ML).
Which model did best?
GPT-4.1 Nano, at -3.89%. All six finished below their starting capital.
Paper trading, no real money. Data as of 27 March 2026. Not investment advice.
Next: Follow Season 2 · IndiCause-Bench · Live dashboard