What we've found so far
Every test on this site in one place: what held up, what didn't, and what we would test next. Updated whenever a new test is published.
Variants tried: 404 · Survivors: 2 · Markets: 5 · Tests: 59
01The short version
- 2 of 404 variants passed every gate, and 1 more passed and was later made Inconclusive by a data check. We have run 59 tests of popular trading ideas across ES and NQ futures, spot gold, four currency pairs and Bitcoin, each in several versions (the FX study re-runs the six ES strategies on FX), under one fixed protocol with every variant reported. 2 variants have earned the label “survivor”, all on one market and on data that earlier tests had used; see the points on the survivors below.
- The development t-statistic ended almost everything. 393 of 404 failed it; the 11 that cleared it were Gold Opening-Range Breakout AS-B, Golden Cross (1H) XAU-GH-L-N, EMA 9/20 gold (M15, H1) GX-H1-L-N, EMA 9/20 gold (M15, H1) GX-H1-L-S, EMA 9/20 gold (M15, H1) GX-H1-L-T3, Noise area breakout (NQ) NY-LS, Noise area breakout (NQ) NY-L, TTM Squeeze (gold, 1-hour) SQ-L-B, WaveTrend (gold, 1-hour) WT-L-R, Heikin Ashi (Bitcoin, 4-hour) HA-NOWICK and Heikin Ashi (Bitcoin, 4-hour) HA-NOWICK-200; 3 of them went on to pass every other gate.
- Costs took back the gross profits. In our first ES study (six strategies) the 28 variants made +$1,264,564 before costs and -$3,548 after, across 41,226 trades. On FX the 96 variants made +125,671 pips before costs and -238,575 after.
- The best-looking ES results are mostly the market going up. In that first ES study the short side lost money in 14 of the 14 variants that trade both ways. Against controls matched on the signal's own entry state or holding pattern, the other long-only winners do not stay above the 95th percentile under every one (their matched percentiles run from 48.8th to 98.6th). One does: RSI Divergence, which still failed the development t-statistic (1.65) and has only 281 trades.
- Hints did not travel. Bollinger's long side was our best hint on ES; on four currency pairs it finished above zero in 0 of its 16 results.
- Our first survivor became Inconclusive. The 1-hour golden cross on gold (long only, no stop) passed every gate, but a later audit found thin, gappy data in its development years; without the trades that touch those gaps its development t-statistic falls below 2.5, so by our data rule the verdict is now Inconclusive. Its forward test continues.
- Two survivors on NQ, and only there. The noise area breakout on NQ futures in the New York session passed every gate in both versions: long and short (1,504 trades, $231,557 per contract after costs, 2020 to 2026) and long only (747 trades, $152,227). The same rule did not pass on ES or on gold, the Asia and London sessions failed on all three markets, and the NQ holdout had been used by earlier tests. They are candidates for a test on new data, not results to trade.
- Some ideas cannot be decided on one market. The 50/200 golden cross gave 3 trades on ES and 15 on gold with the usual long-only rule, so 7 of its 8 results are marked Inconclusive rather than Buried. Read the golden cross test.
This is not proof that no edge exists. “No proven edge” means a statistically convincing edge was not demonstrated by these tests, on this data, with these rules and costs.
02Test by test
By market: futures 148 variants (2 passed), metals 125 (0 passed), FX 151 (0 passed), crypto 19 (0 passed). The market pages list the same tests by market: futures, metals, FX, crypto.
03What ended the 404 variants
| Gate | Failed by |
|---|---|
| 1. PF ≥ 1.2 in validation | 359 of 404 |
| 2. ≥ 50 trades in each period | 32 of 404 |
| 3. Development t-statistic (2.0, or 2.5 where more than four variants were tried) | 393 of 404 |
| 4. Above the 95th percentile of the random control | 307 of 404 |
| 5. Positive without the 10 best trades | 346 of 404 |
| 6. Holdout net positive (already used) | 253 of 404 |
A variant can fail several gates. The random-entry gate uses the control frozen in each specification (for the first ES study and the gold opening-range test, the original control; for FX, the entry-state-matched one; for later tests, the control fixed in their own specification). The first ES posts and the gold opening-range post also show the matched-control check added after review.
04Costs
| Market | Variants | Trades | Before costs | After costs | Cost assumed |
|---|---|---|---|---|---|
| ES futures, first study (six strategies) | 28 | 41,226 | +$1,264,564 | -$3,548 | about $31 a trade |
| Spot gold, opening-range breakout | 12 | 52,616 | +4,774 R | -2,003 R | $0.30 an ounce (a cost check at about $0.17 is in that post) |
| FX majors | 96 | 347,256 | +125,671 pips | -238,575 pips | about 1 pip a trade (1.2 on USDJPY) |
Costs across every test on the site, by market and by how often a strategy trades: How much do trading costs eat? All our backtests compared.
Every test grouped by style: trend following vs mean reversion across all our backtests.
The same entries with and without a stop loss, pair by pair: do stop losses work? 183 backtest pairs.
Most of these strategies trade often. The signals earned something before costs in many versions, and costs took it back: on FX, costs turned 35 of the 65 results that were positive before costs into losses. Costs alone do not explain the failures, though: the results that stayed above zero after costs failed other gates.
05Longs and shorts
In our first ES study, the short side lost money in 14 of 14 long-and-short variants that record the split, while the long side carried the profits that looked good, in a market that rose over the period. On FX there is no drift to lean on, and both sides lost: across the 48 long-and-short variant-pair results the long side made -78,356 pips and the short side -96,197 pips after costs. In spot gold the shorts lost money after costs in 49 of 54 long-and-short variants across our gold tests (the exceptions: LD-A, LD-B, HS-LS-3, SX-STRICT-T1, SX-NEAR-T1). Later ES tests are mixed on this point; each post shows its own split.
06Do the ES winners beat a matched random control?
The long-only versions with the simple exit were the best ES results. The table compares each with the random control frozen before the run and with the controls matched on the signal's own entry state or holding pattern, which we added after an outside review showed the first controls were not always comparable. Percentile of the average net per trade among 500 random runs, 95 needed.
| Test (variant) | Net on ES | Frozen control | Matched control |
|---|---|---|---|
| RSI(2) Dip Buyer (B) | +$26,102 | 91.6th | 90.2th to 98.6th |
| Supertrend (ST60-L) | +$99,304 | 83.0th | 48.8th |
| Bollinger Band Reversal (BB-L-S) | +$103,281 | 90.6th | 81.4th |
| Donchian Channel Breakout (DC-L-S) | +$40,558 | 55.2th | 71.6th to 78.4th |
| RSI Divergence (DV-L-S) | +$48,731 | 78.2th | 98.8th to 99.2th |
| MACD Crossover (MC-L-S) | +$64,167 | 63.4th | 52.8th |
Only RSI Divergence stays above the 95th percentile under every matched control. Its development t-statistic is 1.65 against 2.0 required, it has only 281 trades, and the ES holdout is already used, so it is a hypothesis for a clean test, not a result.
07Did the ES results travel to FX?
| Strategy | Best ES variant | Its net on ES | FX results positive after costs | FX results passing every gate |
|---|---|---|---|---|
| RSI(2) Dip Buyer | B | +$26,102 | 3 of 16 | 0 of 16 |
| Supertrend | ST60-L | +$99,304 | 9 of 16 | 0 of 16 |
| Donchian Channel Breakout | DC-L-S | +$40,558 | 6 of 16 | 0 of 16 |
| Bollinger Band Reversal | BB-L-S | +$103,281 | 0 of 16 | 0 of 16 |
| RSI Divergence | DV-L-S | +$48,731 | 6 of 16 | 0 of 16 |
| MACD Crossover | MC-L-S | +$64,167 | 6 of 16 | 0 of 16 |
Every rule we ran on more than one market, compared version by version: does a strategy work on other markets? The cross-market scoreboard.
Nothing travelled in the sense the protocol asks for: no variant passed every gate on any of the four pairs, so none met the bar of a pass on at least 3 of 4. That bar is the FX study's, not a universal definition of an edge. An edge can exist in one market only; the six strategies on ES and on FX found none in either.
08What we have not tested, and what we would test next
- More markets. YM and crude (CL) futures need 1-minute history exported from the trading platform; gold futures (GC) appear in one test so far (the EMA 9/21 crossover). The 50/200 golden cross is done on ES and gold, but one market gives too few trades to decide it, so the next step there is a longer index history or a test pooled across several markets.
- Other settings and ideas. We have tested 1-minute to daily bars, session rules and a few filters, mostly at default settings, on purpose. We do not tune settings, and news rules are not tested.
- Another hypothesis worth a clean test: long-only RSI divergence, declared in advance with a matched control, on data we have not used. It stands out against matched random entries but is not significant yet.
- A clean test of the NQ survivors: the noise area breakout in the New York session, exactly as frozen, on NQ data that did not exist when it was tested (a forward sample), with the same costs. Until then it stays a candidate. The forward-test protocol (start 12 October 2026) and monthly journal are published.
- A clean test of the gold golden cross: the 1-hour golden cross, long only, no stop, declared in advance again on data we have not used (a forward sample) and on other markets, with real gold costs and swaps. The forward-test protocol and weekly journal are published.
09How to read these results
- Every test follows the protocol: rules frozen and dated before the data, costs included, a random control, every variant reported, and the number of tries counted.
- A holdout period is flagged as already used whenever an earlier test had used the same data, which is true of every test except the first Bitcoin test. ES, NQ and GC are futures data; gold and FX are broker spot (demo) history and Bitcoin is Binance spot history, whose fills and costs differ from live trading.
- Where a post was corrected after review, the correction is logged on that post. Nothing here is financial advice.
10Corrections
15 Oct 2026: the summary gave a number of “popular ideas” (31, then 32) that counted every test as a separate idea, although several tests apply the same idea to another market (the noise area breakout on four markets, the golden cross and the Turtle rules on three); the number was removed. 14 Oct 2026, after a site audit: several statements on this page had gone out of date or were wider than their data. The scoreboard said 3 markets (there are 4, with crypto); the ES cost and short-side figures covered only our first ES study and the gold cost row only the opening-range test, but read as if they covered every test (now labelled); the gold short side was described as flipping between sessions (across all our gold tests the shorts lost after costs in most variants; now computed); and the “not tested” list said we had tested only hourly bars, no session rules and no gold futures, which had stopped being true. No test result changed.
None yet.