De Prado methodology · Edge validation

We close strategies honestly.

Most "profitable" backtests are multiple-testing artifacts. We test not "is PnL positive" but whether the edge survives correction for the number of trials and out-of-sample splits — the Marcos López de Prado standard.

Research archive → Validate my strategy →
Scorecard7 / 7

Every strategy runs seven tests: CPCV (purged + embargo), Deflated Sharpe (n_trials correction), PBO via CSCV, MinBTL, factor-alpha, multi-regime and walk-forward. Passing threshold ≈ 6–7 of 7. We always report the number of trials: a "good backtest" ≠ a discovery.

What's in the archive

18 closed tracks

Public — strategies we tested and honestly closed. It is a showcase of the methodology: what a disciplined null looks like and why most ideas do not survive.

01Ride the Whale — detecting accumulation in the book detect a large trader building a position and step in ahead of them 164,902 events, MFE/MAE ≈ 1.0 (pure noise) Closed 02Hyperliquid — an edge from public on-chain IDs the one venue where participant positions, fills and addresses are public on-chain; does a visible ID solve the problem that killed the tape whale-detector 3 of 4 angles null; the 4th is our existing Bybit spike edge, not new alpha Closed 03Informed Flow Detection volume passes without moving price — someone is absorbing or accumulating, and price follows them 2,090 signals, MFE/MAE = 0.99 across all types Closed 047 Funding Exploits (on majors) systematically sweep 7 mechanics around funding/settlement looking for a structural edge 0 / 7 Closed 05Spread Trading — Calendar / Basis / Cross-Asset trade perp-vs-quarterly spreads, the basis premium, and cointegrated pairs MMs hold the spread <1 bps; no pair is stably cointegrated Closed 06Astrology / Mercury Retrograde vs BTC lunar phases, Mercury retrograde and zodiac transits predict BTC zero edge across all 12 signs Closed 07Icebreaker — breakout momentum momentum at the breakout-bar close (mom-k0) selects continuation of the move 7/7 scorecard turned out to be a 1-bar lookahead artifact (causally 3/7, CPCV −2.48) Closed 08Arbitrage (6 tracks) classic arbitrage: cross-exchange, CEX-DEX, on-chain pools, etc. 6 tracks tested — all dead Closed 09CVD / buy-sell classification (TIB·VIB·DIB·CVD) cumulative volume delta and buy/sell imbalance predict direction 461 signals, all |corr|<0.05, ML acc 47.5% (worse than random) Closed 10Cascade Tick Bars tick/imbalance bars catch the start of a liquidation cascade for a directional entry 3 days: WR 100% (false); 90 days: WR 47%, net −0.10%/trade, Sharpe −3.4 Closed 11TradFi Tick Shadow (NQ → BTC) after a strong NQ (Nasdaq) day, BTC catches up post-close 44 events / 180 days, follow rate 48% (random) Closed 12Signal Denoising (Wavelet + Kalman) denoising price (wavelet/Kalman) improves the directional entry standalone 0/8 validation (Kalman velocity is useful as a tool) Closed 13Denoised Big Move catch large moves on a denoised signal 0/8 validation — long-only beta on a bull market Closed 14Equity Market / PEAD / Micro-Cap port equity-market edges (post-earnings drift, microcaps) onto our stack PEAD ~0.1%/trade (negligible); US 8-K: costs = edge Closed 15News Event Trading (Strategy/Saylor BTC sale) trade the predictable reaction to major news events (BTC sales, etc.) N=2 in all history — not systematizable Closed 16Thin-coin range-MM — “buy the low of every candle” passive market-making on thin coins: buy the low and sell the high of the range on almost every candle, harvest by frequency, gated by regime honest MM sim net −9…−15 bps even at maker=0; 6 independent angles converge Closed 17HL MM-behavior prediction — forced unwind after a liq flush predict a market maker's forced unwind after it absorbs a liquidation flush — MM inventory on Hyperliquid is public on-chain causally-clean NULL across 40 coins / 63 days; partial-corr ≈ 0, realizable fade 0/40 coins Closed 18DeFi liq-map (Aave) — lending liquidations vs a CEX spike Aave v3 LiquidationCall events predict a CEX spike — we read contract state and event topics, not raw network load reverse causation: the liquidation LAGS the crash (pre −110 bps ≈ the whole −116 drop); realizable fade +4.7 bps = noise Closed

Working strategies (LIVE/paper) are in the private section, access on request.

Service: validating your strategy

Send a rule/backtest — we run it through the same scorecard (DSR · PBO · CPCV · MinBTL · multi-regime · walk-forward) with a realistic cost model and tape cross-check, and give an honest verdict: real / artifact / conditional — with numbers and caveats.

What you get
a de Prado-format report: scorecard, cost/capacity, failure-modes, verdict
What we need
a description of the rule + data/backtest (or access to it)
Format
a private report (like this archive, but on your strategy)
Message on Telegram →

Your request goes to our Telegram.