Longbow Weekly PAPER

Candidate: hold-out did not confirm profit

The sealed hold-out (2025-10 to 2026-10-16, opened once on 2026-10-19) did not confirm a profit: -1.4% at standard costs against -3.3% for the Nifty 500, with a 13.8% worst fall. It protected capital in a down year but did not make money. The forward paper run continues.

• Scope. Of your 14 ideas, ten were tested on weekly bars (momentum and 52-week highs, low volatility, trend following, a price-reaction stand-in for post-earnings drift, base breakouts, market structure, EMA support, anchored VWAP, fair-value gaps, RSI/MACD), plus the index-rebalancing event study and pairs. Four could not be tested honestly: quality and value factors (no fundamentals), true post-earnings drift (no results dates or earnings surprises), event-driven and merger arbitrage (no event list), and volatility dispersion (no stock option chains).
• Method. Nifty 500 point-in-time membership, jump-cleaned weekly prices, liquidity of at least ₹5 crore a day, decisions on Friday close and trades at the next open, delivery costs plus 0.2% slippage a side, stress at 1x, 2x, 3x, and a survivorship haircut of 0%, 4% and 8% a year because only 68% to 92% of past members have price history. Ranked on 2016 to 2021 and checked on 2022 to 2025-09. The hold-out from 2025-10 is sealed.
• Screen of 597 variant and market-filter combinations: CAGR above 10% and drawdown under 20% at 2x costs plus the 4% haircut on both periods. Only 1 combination passed in round 1. Every other family failed: best raw results of 15% to 30% a year came with 33% to 53% drawdowns, and base breakouts, low volatility and the RSI/MACD ideas made little or lost money.
• Round 2 (288 focused momentum combinations: ranking, number of names, market filter, stock stop, volatility scaling). 25 passed both periods. The plateau is risk-adjusted momentum (39 or 52 weeks, skip 4) with 10 to 20 names and the market filter on the 26- or 30-week average. Stops and volatility scaling added nothing, so the final rule has neither.
• Candidate (risk-adjusted 39-week momentum, top 15, monthly, Nifty 500 above its 26-week average, otherwise cash), at 2x costs and a 4% haircut: 24.8% a year, worst drop 16.7%, Sharpe 1.40. Train 25.3% and 16.7%; validation 23.3% and 15.3%. Random top-15 picks with the same filter made -0.1% a year, so the ranking, not the filter, carries the skill.
• Stress. At 1x costs with no haircut 32.4% and 13.8%. At 3x costs plus 4% haircut 20.5% and 21.9% (above the 20% limit). At 3x plus 8% 17.2% and 24.6%. At 5x plus 8% 8.3% and 38.0%.
• Weak spots. Calendar years: 2021 +128% and 2023 +81%, with 2018 -10.5%, 2022 -9.3% and 2019 +1.3%. Starting in 2017 or later the worst drop is 22%. Walk-forward (re-pick each year from earlier data): 22.8% a year with a 23.3% worst drop. White reality check over the 288 round-2 variants on the validation period: p = 0.047 (best t-statistic 2.85), borderline.
• Index rebalancing (Nifty 500 additions and deletions, 2016 to 2025-09, about 233 additions and 127 deletions). Buying additions 28 days before the effective date and selling on it gained 0.5% over the index (t 0.7); the larger run-up from 42 days before mostly happens before NSE announces. After the effective date, additions and deletions both beat the index by 1.4% to 1.7% over five days (unclustered t about 3 to 4). Net of about 0.8% round-trip costs and with few events a year, that is roughly 2% to 3% a year at best: not a stand-alone strategy.
• Pairs. Weekly Engle-Granger tests on 300 highly correlated liquid pairs (formation to 2019) found 70 cointegrated; traded 2020 to 2025-09 the 40 tested had a median total of +5.6% at 1x, +1.9% at 2x and -1.9% at 3x. Your named pairs: ICICIBANK/HDFCBANK +18.7% over 5.75 years at 1x (+9.7% at 3x, drawdown 24% to 27% of notional), TATAMOTORS/M&M -46%, INFY/TCS -11%. No edge after realistic friction; shorting needs futures and margin.
• HOLD-OUT RESULT (opened once, 2026-10-19, on the frozen specification, one continuous run from 2016 with positions carried across): 55 weeks to 2026-10-16. Standard costs (paper settings): -1.4%, worst fall 13.8%. At 2x costs plus a 4% haircut: -7.0% (15.8%). At 3x plus 8%: -12.5% (18.4%). Nifty 500 price index: -3.3% (13.1%). Random top-15 picks with the same filter: -14.1% on average (sd 4.5%); the candidate beat 98% of 40 random runs. The market filter was ON in only 56% of weeks, so the account sat in cash 44% of the time. Reading: the ranking showed skill against random picks and the filter limited the fall, but there was no profit; 55 weeks cannot separate skill from luck either way.
• Forward paper run started 2026-10-19 at the owner's instruction, with the specification frozen. It re-simulates from the start date each evening with the same engine as the back-test (a replay of 2024 to 2025-09 reproduced the research equity to the rupee). The hold-out stays sealed. Paper results use real prices with modelled fills, so they exclude real execution.
• CORRECTION (2026-10-20): the hold-out total depends on the rebalance calendar. The -1.4% above is for the cycle phase set by a 2016 start. The same rule with the four-week cycle shifted by one, two or three weeks gave +5.3%, +0.6%, +1.5% and -1.4% over the same 55 weeks (standard costs), all better than the index's -3.3%. The honest summary is 'between -1.4% and +5.3% depending on phase', which is still not a clear confirmation of profit, but it is less negative than first reported. The forward paper run uses one fixed phase; a four-sleeve version that removes phase luck would be a new strategy version.
• Recommendation. Keep the forward paper run as the only clean out-of-sample test left. Do not use real money on this evidence. Use the decision guide in research/weekly/README.md: judge after 26 and 52 weeks, require drawdown under 20% and a return clearly above cash and the index after costs, and do not change the rules during the run. The full documentation, with the exact rules, pass/hold/stop criteria written before the paper results, and the weekly routine, is at /docs/longbow-weekly.html.