The Double Bottom Under the Microscope

1. Motivation

The double bottom — a "W" shape where price falls to a low, recovers, falls again to a similar low, then rises — is among the most widely taught reversal patterns in technical analysis. It appears in virtually every trading textbook and course, almost always presented as a reliable signal that a downtrend is ending. Yet that confidence rests on remarkably little evidence: most presentations show a handful of hand-picked successful examples and never address how often it works, whether it beats random entry, or whether the "quality" rules traders apply genuinely help.

We set out to measure the behaviour directly and — separately — to test honestly whether any measured tendency translates into an exploitable edge. We treat these as two different questions, because a statistically real pattern and a profitable strategy are not the same thing. MPM Markets operates on a principle of reaction, not prediction: we test whether a setup carries genuine forward information before treating it as tradeable.

Research snapshot: 6 markets, 7 years of data, 13,704 patterns analysed, 1-hour timeframe, 3 entry methods tested, 5 quality filters tested. Main result: no statistically reliable predictive edge.

2. Data and Definitions

Instruments. Six CME-group futures spanning three asset classes: Nasdaq-100 (NQ) and S&P 500 (ES) equity indices; gold (GC), silver (SI), copper (HG) metals; crude oil (CL) energy. Chosen to cross asset classes so that any genuine effect would replicate across unrelated markets, while any market-specific fluke would stand out as isolated.

Sample. 20 April 2019 through 24 May 2026, hourly resolution. Data collected and normalized by the author; timezone handling verified per file.

Pattern definition (deliberately flexible). Rather than a textbook "two equal lows" rule, we detect a family of double-bottom-like structures: two confirmed swing lows (each the lowest point within four bars on either side), whose prices match within the greater of 0.25×ATR or a small fixed points threshold; separated by 8 to 120 hourly bars; with a middle peak (neckline) at least 0.5×ATR above the higher low. This permissiveness lets the data — not our prior beliefs — decide what matters.

Significance. Every effect is tested against a randomized benchmark (entries placed at random in the same market and period, held the same duration). An effect is graded SIGNAL if it beats ≥97.5% of random draws, LEAN if 84–97.5%, and drift if below. Because we tested roughly thirty-five market-by-question combinations, we trust a finding only if it reaches SIGNAL and replicates across multiple markets — a lone significant cell is treated as noise.

3. Principal Finding: The Pattern Does Not Predict

Across all 13,704 patterns, raw forward outcomes were 52.2% positive at 6 hours, 54.0% at 24 hours, and 57.8% at 120 hours. Taken naively this looks bullish — but it is exactly what market drift produces in a rising market, and establishes nothing on its own. The randomized benchmark is what separates signal from drift.

No quality filter carries information. Each filter compares high-quality against low-quality patterns within the same period, a comparison drift cannot distort. Pooled across all markets (n = 13,704) at the 24-hour horizon, four of the five filters — similarity (top 25% vs rest), prior trend down vs other, volume high vs low, and tight vs loose lows — fall into the drift zone with z-scores below ±1.5. Only the neckline-break filter reaches SIGNAL (z = 24.65), and it is examined separately below because the effect proves mechanical.

Notably, prior trend ran opposite to textbook expectations — patterns occurring after downtrends did marginally worse — and volume was significantly negative in gold. None of the traditional quality beliefs survived.

The similarity score has no predictive gradient. We designed a 0–100 similarity score to capture everything that should make a double bottom cleaner. Sorted into deciles by score, the relationship between score and 24-hour return is non-monotonic noise: the highest-scoring decile (3.70 points) is beaten by the seventh decile (5.89 points) and barely differs from the second; middle deciles are negative. The intuition that a prettier pattern is a better pattern is, on this evidence, simply false.

4. The Neckline-Break Illusion

One effect appeared overwhelming. Patterns breaking the neckline within 24 hours outperformed those that did not — SIGNAL in all six markets simultaneously, with extreme margins (z-scores of 17 to 25). Gap sizes were large: ES 39.78 points, NQ 175.11 points, GC 22.12 points, SI 0.51, CL 1.39, HG SIGNAL at z = 20.60.

This is an artifact, not a signal. The measurement window included the break move itself. Saying "patterns that broke upward then showed higher returns" is close to saying "things that went up, went up." It is mechanically true and untradeable, because by the time the break is visible, the move it measures has already occurred.

When the same effect is measured forward from the break — the only part a trader could capture — it collapses everywhere: ES −0.18, NQ −0.21, GC +1.18, SI +0.34, CL −0.53, HG +0.10. The double-bottom neckline break is coincident, not predictive.

5. Tradability Testing: Three Entry Points, All Fail

To be sure the pattern was untradeable rather than merely mis-measured, we tested three distinct entry points, each benchmarked against random entries in the same period. The formation and break tests included patterns that went on to fail (broke downward), so no survivorship bias could flatter the results.

Post-break continuation (enter on the break). Entering at the break bar and measuring forward — the only part a trader can capture — produces no horizon with multi-market SIGNAL. SIGNAL counts across 6-hour / 24-hour / 120-hour horizons were 0/0/1 markets. The lone SIGNAL (HG at 120h, z = 2.38) is isolated, with its shorter horizons flat — the textbook chance result. The break does not lead the move.

Formation entry (buy the W as it forms, including failures). The earliest entry: buy as the second low confirms, before any break, including patterns that later fail. Again, no horizon shows multi-market signal. Crude oil shows weak LEANs across all three horizons — the most coherent flicker in the study — but it is one market of six, never reaches SIGNAL, and involves tiny per-trade magnitudes. The isolated HG-120h cell reappears, reinforcing that it is a quirk of copper's slow drift, not a double-bottom effect.

Across both entry methods and three horizons, the double bottom does not produce a tradable forward edge in any of the six markets.

6. Conclusions

Key findings at a glance: 13,704 patterns tested across six futures markets, three entry methods, and five quality filters; every effect benchmarked against randomized entries (drift-controlled); no statistically reliable predictive edge at any entry point; no quality filter — including a purpose-built similarity score — carried information; the neckline break effect proved mechanical (coincident), not predictive.

The double bottom is descriptive, not predictive. Across six futures markets and 13,704 occurrences, it does not forecast price at any tested entry point, and none of its traditional quality filters add information. The neckline break, the one effect that appears significant, is coincident with the move rather than leading it — a confirmation of price action already underway, not a prediction of price action to come.

7. Practical Takeaway

For a practitioner, the useful conclusion is informational, not directional. The double bottom is a legitimate way to read what the market has done — the neckline is a real reference level around which price reacts — but it is not a reason, by itself, to enter a trade. Don't trade a setup because it is a famous pattern; verify every belief statistically; and beware confirmation that arrives too late. Anyone trading double bottoms profitably is most likely being helped by some other factor they have not isolated. The value of this finding is in understanding market behaviour, not in a ready-made trade.

8. Limitations

Bull-market regime only. 2019–2026 rose overall. The pattern's behaviour in a sustained bear market is untested; a negative result here is strong but not a universal law.

Hourly timeframe and one detection rule. A different timeframe or a materially different definition of "double bottom" could yield different results, though the flexible definition was chosen to minimise this.

Finite entry space. We tested three entry points; other exit models, holding periods, or regime filters could in principle differ. Our negative result applies to what was tested.

Hypothetical. All results are derived from historical data with modelled execution; they are not live trading records.

9. What a Researcher Could Try Next

Different market regimes — particularly a sustained bear market, the one condition this study could not test.

Different timeframes — daily and intraday-minute structures may behave differently from hourly.

Adaptive, volatility-scaled tolerances for pattern detection.

Intraday session effects — whether patterns completing in specific sessions (e.g. the US open) differ.

The crude-oil flicker — the one weakly-consistent cross-horizon result, worth a dedicated, pre-registered test before any weight is placed on it.

Why we publish a study with no edge: most trading research shared publicly reports only successes. We believe rigorous research should also report findings that fail to produce an edge. Publishing negative and null results reduces survivorship bias, guards against overconfidence, and gives a more honest picture of how markets actually behave. This is the essence of reaction, not prediction.

10. Reproducibility

Pivot lookback (swing confirmation): 4 bars each side. ATR length: 14. Low-match tolerance: max(0.25 × ATR, small fixed points per market). Min / max separation between lows: 8 / 120 hourly bars. Min neckline depth: 0.5 × ATR above the higher low.

Forward horizons: 6, 24, 120 hours. Random simulations per test: 1,000–5,000.

Total patterns: 13,704 (ES 1,841; NQ 1,671; GC 2,355; SI 2,723; HG 2,404; CL 2,710).

All figures are hypothetical and derived from historical backtesting, which has inherent limitations (hindsight, no real capital at risk, no full accounting for execution factors). The test period covers a broadly rising market and has not been validated in a sustained downtrend. Past behaviour does not guarantee future results. Methodology and data definitions are stated so that results can be independently reproduced and tested using equivalent data.