All strategies · Longbow Learned review page

Longbow Learned: machine-learning stock ranking, documentation

Version 1.0. Frozen 2026-10-20. Owner: Prashant Maheshwari. Status: forward paper run active (₹20 lakh notional), no real money, no orders. It is an A/B partner of Longbow Weekly (same data, same costs, same portfolio rules). Read research/weekly/STRATEGY.md first: this document only describes what is different.

1. Why this exists

The owner asked whether a TensorTrade-inspired approach was worth building. Review result (see section 8): do not adopt the TensorTrade library; borrow its ideas (separate action / reward / exchange components, costs inside the reward), run the work on Nifty 500 weekly data only, in three phases with stop rules fixed in advance: 1. Phase 1: a plain machine-learning stock ranker versus the hand-written Longbow rule. 2. Phase 2: a small reinforcement-learning agent that only sets market exposure. 3. Phase 3: a forward paper run of whatever survives.

2. The strategy in one paragraph

Every Friday, a LightGBM model (an ensemble of 8 random seeds, retrained every 26 weeks) estimates for each liquid Nifty 500 stock the probability of being in the top 20% of 4-week forward returns. That probability is turned into a rank, averaged 50/50 with the rank of the Longbow score (risk-adjusted 39-week momentum, skipping 4 weeks), and the top 15 are held with the same rules as Longbow Weekly: monthly rebalance, hold while rank is within 22, cash when the Nifty 500 is below its 26-week average, trades at the next open, delivery costs plus 0.2% slippage a side.

3. Exactly what differs from Longbow Weekly

3.1 Features (31, all known at the Friday close)

Stock features (cross-sectional percentile ranks among eligible stocks that week): returns over 1, 2, 4, 8, 13, 26 and 52 weeks; momentum 39-to-4 and 52-to-4 weeks; volatility over 13, 26 and 52 weeks; risk-adjusted momentum (two versions); distance from the 52-week high and the 26-week low; price relative to the 10-, 30- and 40-week averages; ATR as % of price; log traded value; volume relative to its 10-week average; RSI(14); beta to the Nifty 500 and residual volatility (52 weeks). Market-state features (same for all stocks that week): Nifty 500 return over 4, 13 and 26 weeks, whether it is above its 26-week average, its 26-week volatility, and the share of stocks above their 40-week average.

3.2 Label and training

• Label: 1 if the stock's return from the next week's open to the close 4 weeks later is in the top 20% of eligible stocks that week, else 0. (A version that subtracts the index return gives the same ranks, so it is not a separate target.)
• Training rows: eligible stock-weeks up to 5 weeks before the block start (the embargo prevents the 4-week label from using future prices). First block trains on everything up to 5 weeks before the paper start; later blocks use all data up to 5 weeks before their start.
• Model: LightGBM classifier, 250 trees, learning rate 0.03, 8 leaves, at least 300 rows per leaf, row sample 0.7, column sample 0.7, L2 penalty 10, 8 seeds (11 to 18) averaged. These settings were set once and not tuned (sensitivity cells in section 4.3).

4. Evidence (development data only; out of sample means each prediction used a model trained only on earlier data)

Window: predictions from 2018-01 to 2025-09, capital ₹20 lakh, 2x costs plus 4% a year survivorship haircut (same convention as Longbow). The 2025-10 to 2026-10 period was not used for evaluation (it was spent as Longbow's hold-out); it is used only as training data in the live run.

4.1 Phase 1: ranker versus rule (N=15, filter on)

• Six ranker types were tried (three models, two target definitions that turned out identical), each as stand-alone and blended, at 10/15/20 names with and without the filter: 72 combinations (36 distinct). Only the classifier family beat the rule; ridge and the regression model were worse. 33% of combinations beat the rule; the median combination lost 4.1% a year to it.
• Information coefficient (rank correlation between score and next-4-week return, 2018 to 2025-09): rule +0.045 (t 2.9); ridge +0.052; LightGBM regression +0.062 (t 3.8); classifier +0.020 (the classifier targets the top tail, not the whole ranking, so IC understates it).
• Multiple-testing check: White's reality check on "does any ML combination beat the rule": p = 0.072. This is above the 5% bar, so the result is borderline, not proven.
• Most important features by gain: market breadth, 26-week market return, market volatility, risk-adjusted momentum, beta, residual volatility, liquidity.

4.2 Stability of the chosen version

• 8 random seeds: CAGR 27.8% to 32.6%, worst fall 19.2% on average (max 20.3%), advantage over the rule +4.9% to +8.6% a year, positive in all 8 seeds. Seed-average ensemble: 30.8%, 19.8%, advantage +7.2% (t 2.72).
• Calendar years (ML blend / rule / Nifty 500): 2018 -11.6 / -9.9 / -4.4; 2019 -0.3 / -0.5 / +9.4; 2020 +43.8 / +39.1 / +13.9; 2021 +170.2 / +125.1 / +29.6; 2022 -9.4 / -13.6 / +0.7; 2023 +71.0 / +61.2 / +27.2; 2024 +24.9 / +7.8 / +14.6; 2025 (to Sep) +7.2 / +2.4 / 0.0. Better than the rule in 7 of 8 years; like the rule it leans on 2021 and 2023.
• Costs and haircut (ML blend / rule): 1x, 0%: 37.7% / 17.3% vs 28.7% / 19.5%; 2x, 4%: 30.8% / 19.8% vs 22.1% / 22.0%; 3x, 4%: 27.3% / 23.6% vs 18.7% / 24.0%; 3x, 8%: 24.2% / 25.8% vs 15.8% / 25.5%; 5x, 8%: 17.5% / 32.7% vs 9.3% / 29.8%. The worst fall exceeds 20% under harsher stress, as for the rule.
• Start dates (CAGR / worst fall): 2019: 38.6% / 18.8%; 2020: 46.5% / 18.8%; 2021: 45.6% / 18.8%; 2022: 22.5% / 22.2%.
• Rolling 54-week windows: negative in 17% (rule: 26%); worst -16.3% (rule -21.7%); median +32.2% (rule +22.9%).
• Null: rule blended 50/50 with random numbers: 15.0% a year (max 22.4%). So the ML part, not the blending, adds the return.

4.3 Plateau check (3 seeds per cell, all cells reported)

Top-quantile threshold x retrain frequency: advantage over the rule between +6.3% and +10.3% a year in all nine cells (CAGR 28.4% to 32.4%; worst fall 19.0% to 21.3%). The result does not hinge on 20%/26 weeks.

4.4 Phase 2: reinforcement learning (exposure only)

Tabular Q-learning agent choosing 0%, 50% or 100% exposure to the stock sleeve each week from market-state buckets and its current exposure; reward = return net of change costs minus a variance penalty; trained walk-forward by year; 20 seeds; OOS 2020-01 to 2025-09. * Always 100% invested: 37.5% a year, worst fall 32.2%. Plain Nifty 500 above 26-week filter: 39.1%, 19.4%. * RL agent (mean of 20 seeds): 16.7%, worst fall 28.1%, average exposure 49%; behind the plain filter in every seed (by 14.6 to 23.2 points a year). Random exposure policies: 2.4%. * Verdict: stopped. With about 300 training weeks, the agent cannot beat a one-line rule. No RL component is used in Longbow Learned.

5. What was deliberately not done

• Nifty options and CRUDEOILM options: not attempted. Only about 22 months of local Nifty weekly-option data exist; earlier work found no edge after costs; there is no CRUDEOILM option data at all and only about 3.5 years of futures. A learning model on so little data would memorise noise. The right first step, if wanted, is sourcing years of historical option prices.
• Nifty index or futures overlay: the RL test above covers exposure; the plain filter is already in use.
• TensorTrade library: not adopted (single-asset crypto focus, last PyPI release 1.0.3 in 2021 with a beta fork, heavy dependencies; its own README states commissions exceed predicted profits).

6. Paper run mechanics

• Script research/ml/paper_ml.py, cron research/ml/nightly_paper_ml.sh Monday to Friday 19:30 IST (after Longbow Weekly's 19:15 run). Outputs in data/paper/ml_weekly/ (equity.csv, trades.csv, positions.csv, orders_next.csv, signals_latest.csv, events.log, and cached model files in models/).
• Deterministic replay from the start date; completed weeks only; placeholder week for the final decision (same as Longbow Weekly).
• Block 0 is trained on the first completed decision week (2026-10-23) using all data up to 5 weeks before; block 1 on 2027-04-23; and so on. Training takes about one minute on 8 cores.
• First decision week ending 2026-10-23; first simulated fills 2026-10-26 open.
• Plumbing validated on development data only (a replay 2024-06 to 2025-09 trained three blocks and produced consistent equity, positions and orders).

7. How to read it for a decision (written before any paper result)

1. It is an A/B test against Longbow Weekly. Both start the same day with the same data and costs. The question is not "is it profitable?" alone but "does the ML blend beat the plain rule by a margin that survives noise?"
2. Check dates: 26 weeks (week ending about 2027-04-23, 7 rebalances) and 52 weeks (about 2027-10-22, 14 rebalances), same as Longbow Weekly.
3. PASS (permission to consider replacing or adding to the rule in a pilot): after 52 weeks, paper worst fall under 20%, annualised return at least 5 points above cash after costs, and at least 3 points a year ahead of Longbow Weekly with the advantage visible in at least 7 of the 12 months. Both strategies share most holdings, so a small gap means "no reliable difference", which counts as hold, not pass.
4. HOLD: positive but ahead of Longbow by less than 3 points, or ahead on too few months.
5. STOP: behind Longbow Weekly by more than 3 points after 52 weeks, or worst fall above 25% from peak. Then retire the ML layer and keep the rule.
6. Retraining is part of the strategy. Any change to features, model, seeds, threshold or blending weight is a new version and restarts the evidence.

8. Risks specific to this strategy

9. Known issues

• The "excess-over-index" training target produces the same ranks as the raw target (subtracting a weekly constant does not change ranks); only one target is used. Variants labelled "ex" in research/ml/results/portfolios.csv are duplicates.
• The first version of the RL script indexed the wrong array; fixed before any result was used.

10. Where everything is

11. Execution audit (added 2026-10-20)

Longbow Learned uses the same audit design as Longbow Weekly: trade log with signal and fill dates, ranks and prices; fills.csv; the automatic checks; one hash-chained, GitHub-committed signal snapshot per decision week (here with the model probability beside each score); the nightly replay check; and the in-browser hash verifier. See research/weekly/STRATEGY.md section 13 for the full description and its limits. The model files are cached per block in data/paper/ml_weekly/models/.

12. One-time check on the 2025-10 to 2026-10 period (pre-registered 2026-10-20, before it was run)

This period was never used to evaluate Longbow Learned (only Longbow Weekly's hold-out was opened on it). It is run once, before the first live model is trained on it (2026-10-23). * Procedure: the same walk-forward procedure as the development tests (retrain every 26 weeks on data up to 5 weeks before the block, 8 seeds, 50/50 blend with the rule), continued through 2026-10-16; no tuning. * Reported at standard costs (paper settings) and at 2x costs plus the 4% a year allowance; the verdict uses 2x costs plus the 4% allowance (the development convention). Both are shown. * Verdict rules over the 55 weeks (total return, ML blend versus the Longbow rule on the same basis): * Pass: ML blend ahead of the rule by at least 3 percentage points and its worst fall under 20%. Confidence rises; the paper run continues as planned. * No reliable difference: within 3 points either way. The paper run continues; expectations stay modest. * Unproven: more than 3 points behind the rule, or worst fall above 25%. The paper run continues at low priority and is not extended. * The result cannot reopen or alter the frozen specification; the period is then spent for this strategy.

13. Result of the one-time check, 2025-10-01 to 2026-10-16 (run 2026-10-20, rules in section 12)

Nifty 500 price index over the same weeks: -3.3% (worst fall 13.1%). Rule blended with random numbers: -1.5% on average.

Pre-registered verdict: NO RELIABLE DIFFERENCE (within 3 points of the rule). Both were positive in a down market; the ML blend was not better than the plain rule in this year. Monthly returns at standard costs, ML / rule: Oct 2025 2.6 / 0.5; Nov -0.6 / -0.1; Dec 0.1 / -1.8; Jan -6.2 / -3.7; Feb 0.3 / 0.3; Mar 0.3 / 0.3; Apr 0.4 / 0.4; May 0.3 / 0.3; Jun 0.3 / 0.3; Jul -4.3 / -2.3; Aug 11.1 / 9.3; Sep 1.5 / 1.9; Oct 0.2 / 0.2.

Rebalance-phase sensitivity (analysis only, no rule change): shifting the four-week rebalance cycle by 0, 1, 2 or 3 weeks gave hold-out totals (ML / rule, standard costs) of +5.4 / +5.3, -2.6 / +0.6, +0.8 / +1.5 and -0.2 / -1.4 percent. In the development window (2018 to 2025-09, 2x costs + 4% haircut) the ML blend was ahead of the rule in all four phases (30.8 / 22.2, 29.2 / 20.9, 24.3 / 20.2, 29.8 / 27.5 CAGR). So the hold-out year gave no sign that the ML layer helps, and no sign that it hurts; its development advantage is intact but not reinforced.

What this changes: nothing in the frozen specification. Under the written rules the paper run continues with modest expectations. The period is now spent: from 2026-10-23 the live models train on it. The same one-phase caveat applies to the paper run, which uses a single fixed phase.

Research and education only. Hypothetical and back-tested results have inherent limitations and do not guarantee future results. Not investment advice; the publisher is not presented as a SEBI-registered investment adviser or research analyst.