v5.6.7: What We Learned From 4,500 Days of Bitcoin Data

Version 5.6.7 of the BV-7X signal engine is live. It’s the biggest single upgrade since v5.6.6’s regime-adaptive thresholds. Three new data categories feed into the model, the validation framework is stricter than ever, and we spent two days diagnosing every period where the model underperformed.

The headline number: 63.3% walk-forward out-of-sample accuracy across 18 validated folds spanning 2016–2025. Up from 62% in v5.6.6.

That 2.6 percentage point overfit gap is the lowest we’ve ever measured. It means what the model sees in training is close to what it gets in real markets. We’ll explain why below.

---

What Changed

We added three new signal categories to the model. Each one was backtested independently before integration, and each had to pass a strict test: improve accuracy without reducing signal count below the previous version.

1. Macro Liquidity Conditions

The signal engine now reads real-time macro liquidity data — a composite of central bank balance sheet movements, credit spreads, and real interest rates. When multiple liquidity indicators agree on direction, they adjust the model’s confidence in its primary signal.

This isn’t a binary flag. The model tracks how fast conditions are changing over multi-week windows and uses the rate of change, not the level, as the modifier. Levels are noisy. Deltas have signal.

Liquidity loosening doesn’t mean “buy Bitcoin.” It means the model can be more confident in a buy signal it was already generating for other reasons. The distinction matters.

2. Stablecoin Supply Momentum

Stablecoin market cap correlates with BTC price at roughly 95%. That correlation is useless for prediction — it’s just two things moving together. But the rate of change in stablecoin supply tells a different story.

When stablecoin supply is growing faster than its recent trend, it suggests capital is staging on the sidelines. When it’s contracting, capital is leaving. We measure this over a rolling window and use it as a directional confirmation signal.

3. Options-Implied Volatility

This is the newest addition and the one we’re most interested in long-term. The model now ingests BTC implied volatility data from the options market — specifically, how fast implied vol is moving over short windows.

The logic is contrarian: a sharp spike in implied vol means the market is panicking. Historically, panic spikes that coincide with a bullish primary signal tend to resolve upward. Conversely, a collapse in implied vol — complacency — that contradicts a sell signal is a warning.

We have five years of daily options volatility data for backtesting. The effect is small but consistent, and it’s orthogonal to everything else in the model. That’s what matters.

---

Diagnosing the Worst Folds

Walk-forward validation splits the backtest into 18 time windows, each roughly six months. The model trains on everything before the window, then predicts the window blind. Some windows are easy. Some are brutal.

The pattern is stark. The best folds are strong trends. The worst folds are sideways markets where Bitcoin drifts without conviction.

Period	OOS Accuracy	Market Condition
Aug 2020 – Feb 2021	80.2%	Strong trend (euphoria)
Aug 2017 – Jan 2018	75.3%	Strong trend (euphoria)
Sep 2023 – Mar 2024	66.7%	Bull recovery
Mar 2024 – Sep 2024	51.1%	Choppy consolidation
Mar 2023 – Sep 2023	43.8%	Sideways mild bull

The worst period — March to September 2023 — is the canonical “trap zone.” Bitcoin sat between $26K and $30K for six months, technically above its 200-day average but going nowhere. The 7-day base rate was essentially a coin flip: 44% of the time, price was higher a week later.

We tested every approach we could think of to fix these folds: tighter confidence thresholds, volatility filters, range compression detectors, penalty functions for low-conviction regimes. Nothing worked. The features that separate a correct signal from an incorrect one in trending markets are statistically identical in sideways markets.

A model that claims 70%+ accuracy across all market conditions is either lying about its methodology or hasn’t tested on sideways data. Sideways markets are genuinely unpredictable at the 7-day horizon. The correct action is to hold, and our regime classifier does that for most of these periods.

This is why we report the overfit gap alongside accuracy. A 63.3% model with a 2.6pp gap is more trustworthy than a 65% model with a 6pp gap — the second one is seeing ghosts in the training data that won’t repeat.

---

Validation Improvements

v5.6.7 introduces a 7-day purge gap between training and test windows. Previously, the last 7 days of training data had labels that overlapped with the first days of the test window. That’s a subtle form of leakage — the training period “knows” something about the test period’s outcomes. Purging those 7 days removes the leak entirely.

The purge gap reduced our signal count from ~1,435 to ~1,370. It also reduced the overfit gap from ~3.5pp to 2.6pp. That tradeoff is unambiguously worth it: we lost noisy signals and kept clean ones.

Every version of the model going forward will use the purge gap. We can’t un-see the leakage.

---

The Upgrade Path

Version	Key Change	OOS Accuracy
v5.6.5	Baseline (chop filter, momentum)	60.0%
v5.6.6	Regime-adaptive thresholds	62.0%
v5.6.7	Macro liquidity + stablecoin + DVOL	63.3%

Each version adds roughly 1–2 percentage points of genuine out-of-sample improvement. The gains are shrinking — which is expected. The easy improvements are already captured. What remains is hard, orthogonal signal that requires new data sources.

We’re currently collecting daily options skew data that will take several months to accumulate enough history for backtesting. When it’s ready, it becomes the next candidate signal.

---