Chart patterns fail chance test across 18,157 combinations

By Daniel Okafor

A TradingView study of 28 chart patterns across 18,157 market-and-chart combinations found no measurable advantage over chance. The author, publishing under the handle hectorldu, reported that the patterns "as a whole" cannot be told apart from random price movement.

The chart carries the title "28 chart patterns against chance: 18,157 tests" and sits on the Bitcoin / TetherUS chart on Binance, coded BINANCE:BTCUSDT. hectorldu closed the analysis with a clear limit on its use: "Educational content. Not investment advice."

What did hectorldu test?

hectorldu opened the chart description with a question aimed at the core of technical analysis: "Head and shoulders, double tops, triangles, flags, wedges: do they predict anything when they are all measured at once and against chance?"

The search covered 28 chart patterns, arranged in 14 families. He then paired those patterns with 76 market series and 23 configurations drawn from eight chart types, including candles, Heikin Ashi, Renko, Kagi, and point and figure. Together, those inputs produced 18,157 combinations of pattern, market and chart.

The same search was then repeated on random walks carrying the same volatility as the real series, which gives the test its benchmark against chance.

Each chart type changes how the same prices appear. Heikin Ashi smooths candles, Renko builds fixed-size bricks, Kagi reverses lines on set price moves, and point and figure ignores small fluctuations. The 23 configurations therefore test the pattern rules against different renderings of the same series, not only against different markets.

How many combinations survived?

Only 15 of the 18,157 combinations survived, according to hectorldu. Another 292 were classified as doubtful. The remaining 17,850 were discarded.

The screening numbers came before the decisive test, which asked whether the strongest readings on real markets were exceptional once random walks of matched volatility entered the comparison. Fewer than one in 1,000 combinations cleared the initial screen.

How did real data compare with chance?

On real data, the strongest signal registered |z| = 9.0. On random walks, the 95th percentile of the strongest signal reached 26.2.

A z-score standardizes how far an observed result sits from its baseline, and a 95th percentile means the top one random run in twenty reached at least that level. Read together, the two figures show that generated series regularly produced more extreme readings than the real market did in this test.

hectorldu reported that chance produces more striking signals than the market does, with p = 0.47. A multiple-testing correction, which he described as a reality check, pushed the result to p = 0.593.

Both p-values sit far above the conventional thresholds for statistical significance, which supports hectorldu's conclusion that the patterns cannot be told apart from chance as a group.

Why do chart patterns look convincing?

hectorldu points to two mechanisms: the eye finds patterns in any series, and with thousands of combinations tested, some will always have worked in the past.

That second point separates this study from a simple success story. A pattern that looks impressive after thousands of trials needs a baseline showing how often chance produces the same result, and the study supplies that baseline while the random walks produced the more striking headline signal.

hectorldu framed the standard in one line: "The useful question is not 'did it work?' but 'does it work better than on a made-up price?'"

What should traders take from this study?

For market participants who trade published patterns, the study sets a demanding standard. A signal must beat what the same rule produces on a random walk of identical volatility before it counts as evidence.

The 15 surviving combinations are the only candidates left under that standard, and hectorldu's verdict leaves them without demonstrated superiority over generated prices. Until a separate test separates those 15 from chance on fresh data, the strongest chart signals in this study carry no statistical distinction from noise.