The daily MACD trend and RSI turning signal on large US stocks and ETFs: does it call the next move?
When the daily MACD trend and an RSI turning point agree, does the stock or ETF keep moving that way by the next close or ten sessions later, by enough to beat trading costs?
None of the four pre-registered trials passed, so the answer is no edge at the bar the study fixed. The judged discovery sample is 23 names, not 34: ten names (EWZ, GDX, GLD, LQD, SLV, TLT, TQQQ, XLE, XLF and META) have daily data only from 2021-10-25 and IBIT only from 2022-09-08, after the 2021-04-19 discovery cut, and SNDK has no bars in the counted window. None of those eleven late counted names contributes a discovery signal, and SNDK contributes none at all; the 34-name count is the 23 discovery names plus the eleven late names. On those 23 names from 2016 to April 2021, calls had 1,879 signals and made +0.211% net to the next close (month-clustered lower bound +0.058%, stress +0.148%, MACD-only baseline +0.063%) and +1.372% at ten sessions (lower bound +0.644%, stress +1.310%, baseline +1.050%). Both call trials fail only the deflated Sharpe test, at 0.0001 and 0.0073 against 0.80; the deflated Sharpe is judged at the family's 14 effective trials, not at the 2,009 trials run across all studies when it was judged. Puts had 1,817 signals and lost 0.187% to the next close (lower bound -0.422%, stress -0.253%, 1 of 5 years positive) and 1.643% at ten sessions (lower bound -2.508%, stress -1.719%, 0 of 5 years, and worse than the MACD-only baseline by 0.055%); both put trials fail the mean, bound, stress, year and deflated Sharpe checks. Nothing passed discovery, so validation and the holdout were never read; the holdout stays sealed. The study prices the stock or ETF, not the option a ten-day trade would buy, on a fixed survivor list in one rising regime, with closes split-adjusted but not dividend-adjusted.
Backtested by QS Research
Pre-registration
The pre-registration fixed the rule before the data was read: on closed daily bars, MACD(12,26,9) sets the trend and RSI(14) turning at its own 14-day average sets the entry; only the first bar of a run counts. The universe is 35 large US stocks and ETFs, a survivor list, with no new signal until 34 daily bars exist. Four trials were fixed: calls and puts, each held to the next close and to the close ten sessions later, pooled across the names from 2016-01-04 to 2024-10-29. Each trial is judged on at least 30 signals, a positive mean net of two times the liquidity tier's per-side cost (plus borrow on puts), a positive month-clustered one-sided lower bound at 0.05/4, a deflated Sharpe of at least 0.80, at most two losing years with at least half the years positive, a positive mean at double costs, and a positive edge over the same-side MACD-only baseline. The 60/20/20 split is by signal date; the holdout opens only for a trial that passes validation. Seventy earlier looks at the same daily rule are counted in the family, so the deflated Sharpe is judged at N = 14. Amendment 1, approved before the run, replaced the minute-data source with a pinned daily-bar snapshot for all 35 names, split-adjusted and not dividend-adjusted, because the minute store held only 19 of the 35. The discovery sample is 23 names; the other eleven counted names start after the discovery cut (ten on 2021-10-25, IBIT on 2022-09-08), and SNDK has no bars in the counted window.
Registered 2026-11-14Amendments
- 2026-11-14 Amendment 1, approved before any counted data was read: the minute store held only 19 of the 35 names, so the data source became an existing pinned daily-bar snapshot, split-adjusted and not dividend-adjusted, with a check that at least 25 names have 250 bars in the counted window. The pin and the run command were named the same day. No rule, window, trial, cost or judging change. approved: Henry, 2026-11-14, before any counted data was read
- 2026-11-21 Erratum after the counted run, from its own output: the daily-bar snapshot starts in October 2021 for ten of the names and September 2022 for one, after the discovery cut, so discovery was judged on 23 names; one name has no data in the counted window and was not listed as unavailable; the per-name dividend print promised by Amendment 1 was not produced. No rule changed and nothing was re-run. approved: Recorded at the independent review, 2026-11-21; it changes no rule
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 4 trials · 0 passed · judged at N = 14 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| Calls (up signals), held to the next close | FAIL | counted |
| Calls (up signals), held 10 sessions | FAIL | counted |
| Puts (down signals), held to the next close | FAIL | counted |
| Puts (down signals), held 10 sessions | FAIL | counted |
Limitations
- The discovery sample is 23 names, not 34. Ten names (EWZ, GDX, GLD, LQD, SLV, TLT, TQQQ, XLE, XLF and META) have daily data only from 2021-10-25 and IBIT only from 2022-09-08, after the 2021-04-19 discovery cut, and SNDK has no bars in the counted window; the eleven late counted names contribute no discovery signal and SNDK is not counted. The 34-name count is 23 in discovery plus the eleven late names, and the 35-name universe is a survivor list, not the live most-liquid 400.
- The stock or ETF is priced, not the ~10-day option such a signal would trade; the option also pays premium decay and its spread, so a stock result that does not pass does not become a winning option.
- Correlated names (SPY, QQQ, SMH, IWM and the large tech names in discovery) move together; the month cluster absorbs some of that, not all.
- One regime: discovery is 2016 to April 2021, a rising market for these names. Calls clear the mean, bound, stress, year and baseline checks and fail only the deflated Sharpe.
- Fills are at the close of a closed daily bar; the orders such a signal would prompt come later.
- No dividend adjustment: a put held over an ex-dividend day gains the dividend and a call loses it; this matters most for TLT, LQD, XLF and XLE. The per-name size print promised in Amendment 1 was not produced, so the size is not shown.
- Puts pay the large-cap borrow rate (0.3% a year) for every name, not a name-specific borrow rate; all 35 are large and liquid, so the amount is small.
- The trial count across all studies was 2,009 when this one was judged; the published count is 2,965, the total when the report was assembled. The deflated Sharpe uses the family's 14 effective trials, not either global count.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
2 published versions · current 83c7c95d344c2482, published 2026-11-21