← ClaudeAtlas

pattern-studylisted

Test candlestick patterns, calendar effects and price cycles against forward returns, with multiple-testing correction and date-clustered inference. Use when asked whether chart patterns work, or to analyse patterns and cycles.
saatwik-1157/trade-research · ★ 0 · Testing & QA · score 65
Install: claude install-skill saatwik-1157/trade-research
# Pattern and cycle event study ```bash python tools/patterns.py --universe tools/universe.txt --years 8 --cycles python tools/patterns.py --tickers "EURUSD=X,GBPUSD=X" --years 10 ``` Detecting a pattern is trivial. Whether anything follows it is the only question that matters, and this measures that. ## Three corrections, all load-bearing **Multiple testing.** 23 patterns times 3 horizons is 69 hypotheses; about 3.5 are expected to clear p<0.05 with nothing happening. Benjamini-Hochberg FDR control is applied and both verdicts are printed, so the gap between "naive significant" and "survives FDR" is visible. **Date clustering.** Pooling correlated names treats 35 tickers signalling on one day as 35 independent observations. They are closer to one. Inference is clustered on date; the naive t is printed beside it to show the inflation. **Sign flips.** Two means are reported: pooled (equal weight per observation) and per-date (equal weight per day). When they disagree in sign the row is marked `SIGN FLIP`, and the apparent edge lives in a few crowded dates - one market-wide move counted many times - not in the pattern. ## Reporting Quote the FDR line, not the top of the table. A row marked `naive only` did not survive correction and must not be described as working. If nothing survives, say nothing survived — that is the result, not a failure of the run. Spectral peaks are reported with their share of variance because a random walk produces peaks too. A peak explaining