The Connors pullback book on US stocks: RSI(2), double 7s and cumulative RSI, 2016 to May 2022
In an uptrend, does buying a US stock after a short-term oversold close and selling on the bounce beat simply holding the S&P 500 after costs from 2016 to May 2022?
No. All three pre-registered rules failed to beat simply holding the S&P 500 after costs over 2016 to May 2022. The primary rule (buy when the 2-day RSI is below 10 in an uptrend, sell when the close is back above its 5-day average) made +2.7% a year after costs (+9.5% before costs), with a 61.6% drawdown, against the index's +11.9% a year (price only); its 3,293 round trips averaged +0.09% after costs (+0.28% before). Double 7s did best: +14.1% a year after costs (+17.2% before), a 48.9% drawdown and +133.3% in total against the index's +105.4%, but its +0.30% a month excess is concentrated in two post-crash months (April 2020 +26.1 points, November 2020 +16.5) and the first half lost 0.63% a month, so it fails. Cumulative RSI made +7.1% a year after costs (+13.1% before) with a 48.9% drawdown and fails. The three rules won 65-68% of their trades before costs, but the average loss was about 1.5 times the average win and the worst trades lost 64-70%. All three also fail a Sharpe test that allows for the 1,741 effective trials already in the shared daily record; that is a stricter bar than the pre-registered family of 3 and is disclosed. The family is closed on this history; the holdout stays sealed.
Backtested by QS Research
Pre-registration
Three rules were fixed before any run, on the common-stock universe with market value at least $1B, 50-day average dollar volume at least $20M and price at least $5: (1) close above its 200-day average and 2-day RSI below 10, sell on a close above the 5-day average; (2) close at a 7-session low above the 200-day average, sell at a 7-session high; (3) 2-day RSI summed over two sessions below 35 above the 200-day average, sell when 2-day RSI is above 65. Equal weight, at most 10 names, one unit, long only, no stop (a 1,000x average-true-range stop that never triggers). Discovery is 2016-01-04 to 2022-05-31. The 200-day average has no history before 2016, so the book holds nothing for the first nine months: 77 calendar months, about 68 with holdings. Entries and exits fill at the next open; idle cash earns 0%; the holdout stays sealed. The pass bar is the standard promotion gate: average monthly excess over the S&P 500 after costs, both halves, the single best month removed, double costs, and a deflated Sharpe (a Sharpe ratio corrected for the number of trials already tried) at the counted family. The pre-registration fixed a family of 3; the engine instead counted the effective trials already in the shared daily record (1,632 in the earlier unpinned run, pass 2; 1,741 in the pinned re-run), a stricter bar, and the results disclose it.
Registered 2026-11-03Amendments
- 2026-11-21 Before the run: the three frozen rules are re-run, on the same window and gate, on pinned data snapshots of the daily bars and the monthly panel, from a signed release, and that run becomes the binding result whatever it shows. The earlier panel file had been rebuilt in place, so this is a new run on fixed inputs, not a byte replay. It reproduced every return, check and backtest id; only the deflated Sharpe's trial count changed. approved: Committed before the run on the owner's instruction to close the unpinned-data gap; accepted in independent review, 2026-11-22.
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 3 trials · 0 passed · judged at N = 3 trials; 2972 trials run in total
| Trial | Verdict | |
|---|---|---|
| cp-rsi2 | FAIL | counted |
| cp-double7 | FAIL | counted |
| cp-cum-rsi2 | FAIL | counted |
| cp-rsi2 | FAIL | not counted |
| cp-double7 | FAIL | not counted |
| cp-cum-rsi2 | FAIL | not counted |
Limitations
- Prices are split-adjusted, not total-return: dividends are excluded, which understates a long book.
- The stock universe is a survivor list: it holds the names known when it was built, so a dip-buying book never buys a name that later delisted. This flatters the result.
- The binding run is the 22 Nov re-run on pinned data snapshots (daily bars and the monthly panel, each by content hash). It reproduced the earlier unpinned run exactly: every return, every check and every backtest id.
- Fills are at the next open after the signal close; the overnight gap can go either way and is not corrected.
- Idle cash earns 0%, which understates a book that is often partly in cash.
- One regime: 2016 to May 2022 is a single decade with two fast crashes and the 2022 bear. A no-stop pullback book is short volatility in spirit, so its worst months matter more than its hit rate.
- The 77-month window includes nine months before the 200-day average has any history, when the book holds nothing. The gate counts those months as flat against the S&P 500. On the 68 months with positions, the double-7s 'best month removed' check is +0.07% a month, not -0.04%; its first-half check is still negative (-0.38% on positions, -0.63% on all 77).
- The win rates (65-68%) and average round trips (+0.28%, +0.83%, +0.39%) in the results are before costs. After costs the win rates are 61.4%, 65.9% and 62.7%, and the average net round trips are +0.09%, +0.64% and +0.20%.
- The Sharpe test was counted against 1,741 effective trials already in the shared daily record (1,632 in the earlier unpinned run), not the pre-registered family of 3. That makes the bar stricter, not looser, and every trial already fails a non-Sharpe check.
- The report's S&P 500 series is monthly, so its 20.0% maximum drawdown is not on the same daily marks as the strategies' 49-62%; the two are not directly comparable.
- No stop: the 1,000x average-true-range stop never triggers; a real tail loss is not truncated. The worst single trades lost 64-70%.
- Double 7s beat the index on total return (+133.3% vs +105.4%) but its excess is two post-crash months; that is not a repeatable premium and it fails the first-half check.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
3 published versions · current d96f62047af500ac, published 2026-11-22