Not supported · Instruments & watchlist · 2026-09-30 (basket) / 2026-10-10 (out-of-sample check); timeframe gap fill 2026-10-11
Does picking the best-backtesting stocks give a better live basket?
What this means
Picking the stocks with the best backtests does not pick the best stocks for the next period. When the same rule was applied to data before 2021 or 2023, the picks had averaged +0.75 to +1.53R per setup in the selection period, then did about the same as the whole universe afterwards (differences −0.14 to +0.17R, all ranges spanning zero). Their high backtest numbers were mostly good luck that did not repeat. On this evidence, the live basket's 63-86% backtest win rates are not a guide to its future; universe-level results are the realistic benchmark. The 15m test gives the same answer: the 50 best 15m stocks up to March 2025 averaged +0.68R per setup, then earned nothing special afterwards.
Explanation checked against the data.
By timeframe
| Timeframe | Verdict | n | Result |
|---|---|---|---|
| 15-minute | Not supported | 785 | top 50 picked before 2025-03: +0.68R in-sample, then +0.009R vs universe +0.036R; diff -0.027 (-0.113, +0.062), q=0.74 gap fill G2, 2026-10-11 (trade logs) |
| 1-hour | Not supported | +0.04 (-0.07, +0.14) / +0.02 out of sample | |
| 4-hour | Not supported | -0.05 / +0.17, CIs span 0 | |
| Daily | Not supported | -0.14 / -0.06, CIs span 0 |
- Hypothesis
- The top-N instruments by backtest net PnL keep outperforming the universe.
- Method
- Live basket = top 50 (1h), 20 (4h), 20 (1d) by net PnL % in the filter-off 2016-26 TradingView sweeps (rule reproduces all three lists). Pseudo out-of-sample: same rule applied before 2021 or 2023, scored afterwards vs the universe, cluster CI.
- Result
- In-sample basket: 1h 63.0% win, +0.64R (PF 2.74); 4h 75.9%, +1.03R; 1d 85.7%, +1.30R. Out of sample, top-50 1h minus universe: +0.04 (-0.07, +0.14) [pick <2021], +0.02 (-0.08, +0.13) [pick <2023]; 4h -0.05 / +0.17; 1d -0.14 / -0.06 (all CIs span 0). In-sample means before the cutoff were +0.75 to +1.53R. 15m gap fill (2026-10-11): top 50 by 15m result before 2025-03: +0.68R in-sample, then +0.009R per setup afterwards vs +0.036R for all instruments, diff -0.027 (-0.113, +0.062).
- How much to trust it
- High: TV fills; selection rule reproduced exactly
- Did it change the live rule?
- this rule built the current 90 live alerts (2026-09-30); expect universe-level results going forward
Scripts and data for this study will be linked here when the research code is published on GitHub.
Education and research only. Not financial advice. Backtests use today's index members (survivorship bias).