Why the desk is built this way
Research
Properly controlled studies find no reliable stock-picking skill in LLMs on their own once they test after the model's training cutoff, over long periods, with costs. Rule-based momentum has the better record, and LLMs are useful at spotting red flags. Hence a momentum screen with an LLM veto, next to a control account without it.
Evidence
| Study | Sample | Finding | Verdict | |
|---|---|---|---|---|
| FINSABER (KDD 2026) | US stocks 2004-2024, survivorship-free | Buy-and-hold Sharpe beat the LLM agents FinMem and FinAgent in every setup; no significant alpha. Too cautious in bull years, too aggressive in bear years. | No edge | Source ↗ |
| StockBench | 20 Dow stocks, Mar-Jun 2025, after model cutoffs | Best agent +2.5% vs buy-and-hold +0.4%; every agent trailed buy-and-hold in the downturn; worse with more stocks. | Thin | Source ↗ |
| LiveTradeBench | 21 LLMs traded live for 50 days | General leaderboard rank does not predict trading results. | Thin | Source ↗ |
| KTD-Fin (2026) | China CSI 300, 2024 to Apr 2026, tickers hidden | 9 of 10 agents had negative stock-selection alpha; returns came from market and momentum exposure. | No edge | Source ↗ |
| Live agent fleets (2026) | 7.5M model calls, 14.6k fills | No directional edge; sizing ignored volatility; many winners turned into losses without a mechanical exit. | No edge | Source ↗ |
| Lopez-Lira & Tang | GPT headline scores, 4,123 US stocks | Real but short (1-2 day) drift, mostly from negative news; unprofitable at 20 bps costs and decaying each year. | Narrow | Source ↗ |
| FinCAD (2026) | Memorisation test on backtests | Backtests on dates the model trained on were inflated by up to 67%, most for big famous stocks. | Bias | Source ↗ |
| NSE factor indices | Total return since Apr 2005, partly back-calculated | Nifty200 Momentum 30 about 18.8%/yr and Low Vol 30 about 16.9%/yr vs Nifty 50 about 13.6%/yr; momentum can crash. | Evidence | Source ↗ |
Backtests of this system
| Variant | Universe | Return | Max drawdown | Trades | Charges ₹ | Windows beating Nifty |
|---|---|---|---|---|---|---|
| Momentum screen, swing exits (2xATR stop, 2.5R target) | Nifty 200 | −6.98% | −21.08% | 646 | 12,965 | 1/4 |
| Swing exits + SEPA trend template | Nifty 200 | −10.43% | −28.39% | 721 | 14,584 | 2/4 |
| Slow momentum book (4xATR stop, no target, 20-day hold) | Nifty 200 | −1.50% | −22.60% | 357 | 5,356 | 2/4 |
| Slow book + SEPA (current default)live | Nifty 200 | +10.22% | −20.20% | 473 | 7,199 | 3/4 |
| Technical rules, 8 large caps | 8 Nifty stocks | −7.27% | −14.96% | 475 | 11,004 | 1/4 |
| Nifty 50 buy-and-hold | Index | +4.64% | — | 0 | 0 | - |
Today's index members, so survivorship bias flatters every row. No row is statistically significant.
Strategy change log. Settings (stops, sizing, screen) change only with backtest plus paper evidence, and each change is logged with its reason. Open the log →