An evidence-based evaluation of the signal, risk, and execution tools every trading textbook recommends — backtested out-of-sample on crypto perpetuals and traditional markets with a real cost model, parameter sweeps, regime breakdowns, and multiple-testing correction. Every result was written back to a Notion catalogue; the code, the data pipeline, and the full audit trail live in this repo.
Anyone who has read trading literature meets the same tools over and over: moving-average crossovers, RSI, Bollinger fades, Kelly sizing, risk parity, VIX filters, funding-rate carry. Each has a plausible mechanism — and almost none have clean public evidence. This project took the ~55 tools catalogued in a master Notion spec and ran each one through the same disciplined harness: real data, walk-forward out-of-sample evaluation, costs applied, parameter grids reported as distributions (never best-cell cherry-picking), and a pre-registered pass/fail bar per tool role.
The pattern is stark: classic directional signals (trend, momentum, mean reversion, seasonality, carry) do not produce net-of-cost out-of-sample edge on liquid crypto perps. A few show faint positive information coefficients; none clear the Sharpe bar.
The tools that helped were not alpha generators but risk transforms: exposure caps, correlation-aware netting, and vol-targeted sizing each cut catastrophic drawdowns (−97.6% → −35.3%) with Sharpe/Sortino intact or better.
17 tools (fundamentals, order-book depth, tick flow, options, news) could not be tested with free/available data. Rather than fabricating proxies, each has a documented gap and what would close it — with a few free-but-unbuilt sources built first (CFTC COT).
Every tool ran through the same 7-stage pipeline. Each stage is documented in depth on the Methodology page.
Exposure Caps — 7.8× gross book → capped; OOS drawdown −97.6% → −35.3%, Sharpe unchanged.
Correlation-Aware Netting — same book, netted by correlation; drawdown cut and Sharpe/Sortino improved.
Realized-Vol Targeting — vol-scaled sizing beats fixed size on Sharpe, Sortino and drawdown (relative to a losing base signal).
HMM regime filter — passes on a 1.7pt accuracy edge + OR-bar technicality; gating actually hurt Sharpe.
ATR stops — beat a calibrated fixed-% stop, but both are worse than no stop at all.
Kelly criterion — "validated" only because it trades ~6% of the time.
PCA stress filter — fails classification accuracy, but gating doubled Sharpe (0.32→0.80): a deliberate user override.
Every trend / momentum / mean-reversion signal — IC ≈ 0 or negative, net Sharpe well under 0.3.
Funding-cycle trading — Sharpe −5.46, destroyed purely by transaction costs.
COT positioning, macro filter, implied vol, risk parity, Markowitz MVO — all clean misses.