mykola-quant/turtle-soup-smt-falsification

Turtle Soup and CRT (Canon Range Trading) are names for a popular discretionary trading pattern used by retail traders in forex, metals, and crypto. SMT divergence (Smart Money Technique) refers to a confirmation signal where a correlated instrument, such as gold and the US dollar, fails to move in the expected direction alongside the primary asset. This repository tests whether adding the SMT divergence filter to a Turtle Soup strategy actually improves profitability. The answer, based on four pre-registered tests across gold/silver, gold/dollar, and Bitcoin/Solana pairs, is no. Every version of the strategy produces negative expected returns after realistic transaction costs. Where the SMT filter showed statistical significance in full-sample data (gold versus dollar proxies in 2025), the effect turned out to be concentrated in a specific period and failed to hold across earlier years or a different asset class. The authors call this a regime fingerprint, not a stable edge. What makes this repository notable is the methodology rather than the result. Most retail strategy testing fixes thresholds after seeing the data, charges no transaction costs, and cherry-picks favorable time windows. This project does the opposite: it pre-registers every threshold before running any tests, uses 10,000-shuffle permutation tests for statistical significance, charges realistic costs (0.5 pip round-trip on forex, 0.13% round-trip on crypto), and applies a strict out-of-sample rule where a result survives only if both halves of the data independently show positive returns below a fixed p-value. Three apparent signals that passed full-sample tests were rejected by this out-of-sample requirement. To reproduce the tests, you download 1-minute bar data from HistData (free), run a conversion script to clean it into parquet files, optionally build a synthetic dollar index from six currency pairs, and run the main backtest script. It prints the full result including the permutation test, bootstrap confidence interval, and out-of-sample verdict. Data comes from HistData for forex and metals, Binance for crypto. Built with Python, pandas, numpy, and matplotlib.