We tested 49 machine-learning models on Bitcoin. Zero survived.

Every week someone will sell you a machine-learning model that predicts Bitcoin. We spent months building the experiment that tests whether that product category should exist: a full ML research pipeline over daily BTC data — on-chain activity, derivatives, macro, volatility, positioning — run with the kind of statistical discipline that makes the answer trustworthy either way.

The answer: 49 model-target combinations tested, zero deployable. Not "needs more tuning." Not "promising but early." Zero.

What we actually tested

The pipeline trained models — including LSTM and CNN architectures alongside tree-based learners — against multiple forward-return targets on a feature universe of public daily Bitcoin data. Every combination faced the same gauntlet:

• Walk-forward evaluation — models only ever score data from after their training window, across dozens of rolling splits. No hindsight.
• Realistic costs — roughly 20 basis points per side (≈40 bps per round trip). Most published crypto backtests quietly assume free trading; this alone kills a large fraction of "edges."
• A dual deployment gate — a candidate had to clear economic significance (better than 50 bps/year net of costs) and statistical significance (information-ratio t-stat above 2.0). Both, not either.
• Multiplicity control — the pipeline keeps a ledger of every statistical test it has ever run: 532,780 hypotheses (92,380 in the main ledger, 440,400 upstream). False-discovery-rate control is applied against that entire pool, because testing half a million things and reporting your best one is how most backtests lie. At this scale, the Bonferroni-adjusted per-test alpha is below one in a million.

What happened

After FDR gating, five candidates were promoted to falsification testing — inversion tests, random baselines, parameter perturbation. Four failed outright; one was flagged for review. That final candidate then produced exactly zero trades in forward validation and died at the last gate.

At the final scoring gate, every measured combination — 15 of 15, covering three backtested model families across five return-sign targets — was NO-GO. The remaining 34 of the 49 never earned their way to the final backtest at all, failing by default upstream. Whichever way you count it, the deployable total is zero.

The same verdict repeated everywhere we looked for it: a separate intraday "horse race" of 96 model combinations — null. A pretrained 4.1-million-parameter transformer (2,048-bar context, trained on data from 45+ global exchanges), run against the same 73 walk-forward splits at 1-hour, 4-hour, and 8-hour horizons — null.

The part that makes the null trustworthy

Months after the pipeline was built, we ran a fresh-eyes methodology review on it with one question: did any flaw hide a real edge?

The review found four residual imperfections. Here is the striking part: every single one biases results toward false positives, not away from them. Screening statistics that overstate significance on overlapping windows, entry conventions with no execution lag, test windows that overlap across splits, promotion rankings computed before costs — each flaw makes it easier for a bad model to look good.

A pipeline with those flaws should over-discover. It still discovered nothing. That makes the null conservative-robust: if anything, the truth is more null than reported. (And yes, we publish our own pipeline's flaws. That's the house style.)

Why a null is worth publishing

Because the usual outcome of this experiment is worse than a null. Run 49 combinations without multiplicity control and several will look brilliant by chance; the typical shop deploys the best one, trades it until the money confirms what the statistics would have told them for free, and never publishes anything. A trustworthy null — one that survived costs, walk-forward discipline, falsification, and a half-million-test FDR correction — is rarer and more useful than one more fragile "alpha."

The honest summary of what we learned: predicting daily Bitcoin direction from public daily features with generic machine learning was one of the lowest-prior questions in quant crypto, and the rigorous answer is the one the prior suggested. Meanwhile, everything in our program that did survive research lives elsewhere — in structural premia and regime frameworks, not in a classifier's opinion about tomorrow's candle. Where a genuine informational pulse did appear, it was found by a differently-shaped tool in a timestamped forward trial, not by the model stack.

What we do differently because of this

We stopped. The pipeline is closed, its verdict recorded, and the marginal 50th model-target combination has approximately zero expected value. Reopening it would only spend the multiplicity budget the ledger exists to protect. The daily analysis you see on this site runs on what survived — and every experiment behind it was pre-registered before its data was touched, so the failures stay counted.

That discipline is the product. How the pipeline works →

Research & entertainment only — nothing here is financial advice.