Held-out test
328 days
Autonomous trading agent · simulation only · no real funds
The model has no edge.
We built a Markov regime model, gave it honest rules — daily bars, real costs, data it had never seen — and asked it to beat the dumbest benchmarks in finance. It couldn't. This site is the receipt.
The verdict
It predicts tomorrow’s regime 90.9% of the time. So does guessing “same as today.”
Three predictors, one held-out window. The bars grow to the score each one earned. Watch the top two finish in exactly the same place.
Markov modeltransition matrix
0.0%
Persistencet = t − 1
0.0%
Naivealways majority
0.0%
90.9%
≡
90.9%
No skill added
And not just Bitcoin — re-tested across 20 markets and up to 20 years of history, the edge is 0.000 on every one. Verified pointwise: 0 divergences in 16,773 predictions.
So where does 90.9% come from?
Split the held-out year into the days the regime stayed and the days it actually moved. Persistence scores zero on the movers by definition — so they are the only days where skill could exist.
Stay daysregime unchanged
298 / 298model correct
Switch daysregime moved
0 / 30model correct
The model never once predicted a change. Every point of its accuracy is stay-day credit — which is exactly what persistence is. It isn't useless, though: its probability estimates beat a sticky baseline on log-loss in 28 of 28 runs. It knows something. It just never changes the answer.
How it works · sense → reason → act
01 — Sense
Label the market’s mood.
Every day, fresh bars come in. A volatility-aware z-score sorts the move into one of three regimes — bull, bear, or sideways — so the threshold adapts to the asset instead of a brittle fixed cut-off.
window = 30 · k = 0.2 · z-score02 — Reason
Forecast the next state.
A Markov transition matrix turns history into probabilities: given today’s regime, where does tomorrow land? The thick self-loops say it plainly — markets mostly stay put. That persistence is exactly what the model can’t beat.
3 states · estimated walk-forward · no lookahead03 — Act
Trade it. Pay for it.
Each decision flows through one fill engine with hard risk guards — long or flat only, no shorts, no leverage. Every fill is charged a realistic 0.3% fee + 0.5% slippage. No real funds, ever.
paper ledger · fee 0.3% · slip 0.5%Held-out equity · net of costs
Regime baseline −41.1%
Buy & hold −47.7%
hover, drag, or press ← → · —
| Policy | Return | Sharpe | Max DD | Trades | Cost |
|---|---|---|---|---|---|
| Regime baseline | −41.1% | −1.87 | −42.9% | 6 | 4.8% |
| Buy & hold | −47.7% | −1.17 | −51.2% | 1 | 0.8% |
| LLM agent Phase C | not yet run |
“The transition matrix is adding little beyond regime autocorrelation. High accuracy here is mostly persistence, not skill.”
— the instrument, flagging itself · warnings[0] · results/BTC-USD_4a150b23.json
Disclosures · read before you believe it
01
Reproduced across 20 markets + sub-periods (0 / 20) — and across up to 20 years of history on 8 assets, spanning 2008, COVID and the 2022 bear: 0 divergences in 16,773 predictions total. The identity is not a window artifact.
02
Regimes come from overlapping rolling windows, so consecutive days are highly correlated — which is precisely why persistence is so hard to beat.
03
The test window trended down. A different period could flip the baseline-vs-hold order. Treat the sign as the finding, not the magnitude.