crypto-ml-study

Most retail "crypto ML" work is leakage with extra steps: labels that see the future, cross-validation that shuffles time, and results reported only when they look good. I wanted a research process I could trust — one where a null result is a valid, published outcome.

• A research repo built on the AFML playbook: dollar bars instead of time bars, CUSUM event sampling, triple-barrier labelling with meta-labelling, and PurgedKFold cross-validation with embargo.
• A pre-registration workflow: each study fixes its hypothesis, features, pass bar and evaluation window in a written spec before any model is fitted.
• Held-out, embargoed test windows and explicit leakage checks that every study must report on.
• A ledger of 18+ completed studies, including the ones that failed their pass bar.

• Frozen pipeline pieces are imported, never copied, so a study cannot quietly diverge from the shared data and labelling code.
• Null results are written up with the same care as positive ones — the rigour is the product, not the Sharpe ratio.
• The repository is private because it contains live research; this page is the public write-up.

• 18+ pre-registered studies with documented pass/fail outcomes and leakage sections.
• A reusable, tested feature and labelling pipeline shared across every study.
• Now being extended from spot BTC to Kalshi BTC perpetual markets.

• Extend the study set to Kalshi BTC perpetuals and compare regime behaviour.
• Publish a sanitised subset of the ledger with the pre-registration templates.
• Add walk-forward retraining experiments with explicit drift diagnostics.