Negative results, registered
What we tested
and killed
Every hypothesis below was written down as a spec before it ran: a fixed universe, a fixed cost model, and a fixed pass or fail gate, frozen in a design doc before any result was seen. We then ran it on the same costed engine that produces the live track record, and recorded what the gate said. None passed. Across two final verdicts the standing conclusion is narrow and blunt: single-asset short-term timing is foreclosed once real execution cost is charged, and no candidate that fits a retail crypto cost structure has cleared its gate.
Each entry links its committed machine-readable artifact and the final verdict memo. The running ledger of every backtest is the append-only experiment registry. The live cost-adjusted result is on the track record, and the per-trade cost dataset is served raw at /api/research/cost-field.
Phase 2 · registered 2026-06-10
Single-asset hourly timing
A technical entry on hourly candles (momentum, mean-reversion, or a regime-aware blend) can pick BTC turning points often enough to pay for itself after fees and slippage.
At the live risk geometry with crypto perp costs charged, every entry is net-negative on a trustworthy sample. Strip the cost and three of the five turn gross-positive, so the signals are not random. The roughly 0.4% round trip is simply larger than the edge. The single positive costed cell in the full 108-cell edge-map, classic momentum with wide 4R targets on H1, came to 0.03 to 0.05% per trade: real, and too thin to build on. Final verdict: single-asset OHLCV timing has no deployable edge after costs.
- Universe
- BTCUSD, H1
- Window
- 2024-06-10 to 2026-06-09 (17,497 bars)
- Entries
- classic, regime-aware, hmm-top3, vwap-ema-bb, full-risk
- Geometry
- live ATR14 x2.5, TP 2R
- Costs
- crypto perp, ~0.4% round trip
Live geometry, crypto costs charged
- classic momentum
- −11.1% · 33% wr
- regime-aware
- +0.6% · 50% wr
- hmm-top3 (window-capped)
- −0.2% · 33% wr
- vwap-ema-bb mean-reversion
- −0.8% · 17% wr
- full-risk (window-capped)
- −0.3% · 33% wr
Same entries, cost stripped (legacy 2:1 geometry, zero cost)
- regime-aware
- +1.3% · 62% wr
- hmm-top3
- +0.4% · 67% wr
- full-risk
- +0.3% · 67% wr
- classic momentum
- −4.3% · 32% wr
- vwap-ema-bb
- −0.3% · 17% wr
Phase 4 · registered 2026-06-11
HMM regime routing
If a hidden-Markov model labels the market as trend, volatile, or range in real time, routing each entry only into the regime that suits it recovers an edge the blended signal buries.
The regime model is honest about itself: its states separate volatility, not direction, so there is no trend premium to route toward. On paper the gate fails. The one routed cell with an adequate sample, classic momentum in the trend regime, is negative on all three symbols. Every cell that prints positive is thin, under 30 trades, and the positives disagree across symbols. Not a trustworthy edge.
- Universe
- BTCUSD, ETHUSD, SOLUSD, H1
- Window
- 2024-06-01 to 2026-06-01, walk-forward 4 folds
- Model
- 3-state HMM (trend / volatile / range), trailing-64 Viterbi
- Routing
- {classic, vwap-ema-bb} x {trend, volatile, range}
- Costs
- crypto perp; E = mean pnl% after costs
Only non-thin routed cell, the trend route (classic, n ≥ 142)
- BTCUSD trend route (n142)
- −0.452%
- ETHUSD trend route (n152)
- −0.672%
- SOLUSD trend route (n203)
- −0.190%
Regime model diagnostic (why there is nothing to route toward)
- directional trend premium
- none
- |24-bar fwd return| separation
- 0.0017 to 0.0157
- regime flips per week (mean)
- 8.3
Phase 4.5 · registered 2026-06-12
Daily time-series momentum
The one timing edge the literature says survives real cost is slow: hold the daily trend on a 28-day lookback and ride it to the opposite cross. Test it faithfully across ten majors.
The signal-flip config reads +24.92% on average, but two launch-era single-asset flukes carry it. Strip SOL and AVAX and the typical major loses 5.25%. Four of ten symbols are positive, below the six-of-ten bar, and fold stability is 38% with most fold cells too thin to trust. The geometry-exit variants are flatly negative. No configuration clears the deployable bar.
- Universe
- 10 majors (BTC ETH SOL BNB XRP ADA DOGE DOT LINK AVAX), D1
- History
- ~2,090 to 2,190 daily bars each (2020 to 2026)
- Signal
- 28-day TS momentum; 4 folds; no parameters tuned
- Costs
- crypto perp
- Bar to clear
- mean and expectancy > 0, ≥ 6/10 symbols adequate, fold stability > 50%
Three configs against the deployable bar
- signal-flip mean
- +24.92%
- signal-flip ex-flukes (SOL +189%, AVAX +102%)
- −5.25%
- signal-flip breadth · fold stability
- 4/10 · 38%
- geometry-2R mean
- −12.54% · 0/10
- geometry-4R mean
- −14.01% · 0/10
Phase 5 Track A · registered 2026-06-13
Funding-rate carry
Stop timing price. Harvest the structural funding premium: short the perp, hold the spot, collect the funding, stay delta-neutral. This is the strongest raw crypto edge on record.
The edge is real and low-variance: always-on BTC returns +39.50% over 6.75 unlevered years, with a 0.73% max drawdown and all four folds positive. It fails on magnitude, not existence. Full-window yield is 5.84% per year against an 8% gate, and the recent 24 months pay 2.35% per year, below what a stablecoin returns. The premium has compressed year over year in our own registered data. Adding a timing overlay (A2) or a rotation (A3) makes it strictly worse, because turnover cost eats the harvest.
- Universe
- 10 majors; 7,403 BTC funding events back to 2019-09
- Accounting
- delta-neutral, notional 1 on capital 2, unlevered
- Costs
- two-leg, 0.70% per full round trip; 4 folds
- Gate
- > 8%/yr full-window and > 5%/yr recent-24mo and max DD < 10% and ≥ 3/4 folds +
Three variants, no tuning
- A1 always-on BTC, full-window
- +5.84%/yr (+39.50% over 6.75y)
- A1 recent 24mo · max DD · folds
- +2.35%/yr · 0.73% · 4/4
- A2 threshold-gated (per symbol)
- 0/10 pass
- A3 top-3 rotation, weekly
- −0.07%/yr · recent −6.11%/yr
Phase 5 Track B · registered 2026-06-13
Cross-sectional momentum
Rank the 30-major universe by trailing return each week and hold the top five. Rotation should beat passively holding the same basket, or it is only churn.
The full-window +1325% looks like a win over the basket at +759%, but the whole excess lives in one fold, the 2020 to 2021 launch run measured over today’s surviving 30. Fold stability fails, two of four. In the bias-mitigated subwindow the long-only book returns −46.99% against a basket at −50.29%, only one of four folds positive. The long-short book’s subwindow PASS is reported exactly as the frozen gate computed it, then set aside: a market-neutral book that loses 12% of capital is not deployable, and its real benchmark is cash, not a crashing basket.
- Universe
- 30 majors, D1 (2,190 grid days), listing-date-aware
- Signal
- 14-day lookback, weekly rebalance, top-5
- Costs
- 0.2%/side on actual turnover, charged to the benchmark too
- Gate
- beat equal-weight basket on return and Sharpe, ≥ 3/4 folds of positive excess
Full window
- B1 long-only top-5
- +1325.48% vs basket +758.96% · 2/4
- B2 long-short
- −12.64% vs basket +758.96% · 2/4
Bias-mitigated subwindow (2024-06 onward)
- B1 long-only top-5
- −46.99% vs basket −50.29% · 1/4
- B2 long-short (gate PASS, set aside)
- −12.15% vs basket −50.29% · 4/4
What survived
Nothing did. Five families of short-term edge, each pre-registered and each killed at real cost. The common killer is the cost denominator: roughly 0.4% per round trip against edges that have compressed hard into 2024 to 2026. Funding carry is the one candidate whose raw structural edge genuinely exists, and its blocker is decay rather than cost. The rest are simply too small to survive the fee.
What would change our mind is exact and public: a pre-registered spec that passes its frozen gate on this same costed engine, at real execution cost, with folds that hold out of sample. Until one does, the honest default is to hold, and the durable asset is the bench that keeps killing bad strategies before anyone trades them.