Longbow Learned PAPER

Promising but unproven: hold-out showed no difference

The one-time check on 2025-10 to 2026-10 (run 2026-10-20 under pre-registered rules) found no reliable difference between the machine-learning blend and the plain rule: +5.4% against +5.3% at standard costs (+0.8% against +0.5% at 2x costs plus the 4% allowance), with the index at -3.3%. The development advantage (30.8% against 22.1% a year) is intact but not reinforced. The forward paper run continues.

• Question asked: is a TensorTrade-inspired approach worth building? Answer: not the library (single-asset crypto focus, last PyPI release in 2021, its own README says commissions exceed predicted profits), but its ideas, in a small, time-boxed experiment on Nifty 500 weekly data only, with stop rules fixed in advance. Options (about 22 months of Nifty weekly-option data, no CRUDEOILM option data) were not attempted: too little data for learning models.
• Phase 1. 31 point-in-time features (momentum, volatility, distance from highs and lows, moving-average ratios, ATR, liquidity, volume, RSI, beta, market breadth and volatility). Label: the stock is in the top 20% of 4-week forward returns. Walk-forward: retrain every 26 weeks on data up to 5 weeks earlier (embargo), predictions from 2018-01 to 2025-09. Ridge, LightGBM regression and a LightGBM classifier, alone and blended 50/50 with the rule, at 10/15/20 names with and without the market filter (72 combinations, 36 distinct).
• Result: only the classifier family beat the rule. Classifier + rule blend, 8-seed ensemble: 30.8% a year, worst fall 19.8%, Sharpe 1.51 (rule 22.1%, 22.0%, 1.25; Nifty 500 11.8%, 34.4%). Ridge (11.5%) and LightGBM regression (16.3%) were worse than the rule. 33% of combinations beat the rule; the median combination lost 4.1% a year to it. White reality check p = 0.072: borderline.
• Stability. 8 random seeds: advantage over the rule +4.9% to +8.6% a year, positive in all 8. Nine threshold and retrain-frequency cells: +6.3% to +10.3%. Better than the rule in 7 of 8 calendar years (2021 +170% and 2023 +71% do much of the work, as for the rule). Rule blended with random numbers: 15.0% a year, so the machine-learning part adds the return. Under 3x costs plus an 8% haircut: 24.2% with a 25.8% worst fall (rule 15.8%, 25.5%): the 20% drawdown limit does not survive harsher stress.
• Phase 2. A tabular Q-learning agent chose 0%, 50% or 100% exposure each week (reward after costs, variance penalty), trained walk-forward, 20 seeds, 2020 to 2025-09. It made 16.7% with a 28.1% worst fall, behind the plain Nifty 500 above 26-week average filter (39.1%, 19.4%) in every seed. Random exposure: 2.4%. Stopped: no reinforcement learning is used.
• Phase 3. The blend is frozen and runs as a forward paper run from 2026-10-19 (₹20 lakh notional, no orders), alongside Longbow Weekly as an A/B test. The 2025-10 to 2026-10 period (Longbow's spent hold-out) was not used for evaluation, only as training data.
• Decision rules, fixed in advance: judge at 26 and 52 weeks. Pass (permission to consider a pilot) only if the paper worst fall is under 20%, the return is at least 5 points above cash, and it is at least 3 points a year ahead of Longbow Weekly with the lead visible in 7 of 12 months. Anything smaller is hold. Stop if it is more than 3 points behind Longbow Weekly after 52 weeks or its worst fall exceeds 25%.
• ONE-TIME CHECK (2025-10-01 to 2026-10-16, 55 weeks, run on 2026-10-20 after the rules were committed to GitHub): standard costs ML blend +5.4% (worst fall 10.1%) versus the rule +5.3% (8.0%); at 2x costs plus the 4% allowance +0.8% (12.3%) versus +0.5% (10.5%); Nifty 500 -3.3% (13.1%); rule blended with random numbers -1.5%. Pre-registered verdict: NO RELIABLE DIFFERENCE (within 3 points). Shifting the four-week rebalance cycle by one to three weeks gave ML totals of -2.6% to +5.4% and rule totals of -1.4% to +5.3%, so about 7 points of this year's result is rebalance-calendar luck; in the 2018 to 2025-09 development window the ML blend led the rule in all four phases. Reading: no evidence for or against the learning layer from this year.
• Full documentation (features, training protocol, every table, risks, routine): /docs/longbow-learned.html.