Volatility regime

The same LightGBM as above with the same indicators, but asked a different question: not whether price will rise, but whether it will move more.

Why it matters

It's the most important result in this whole section, and not because of the algorithm — it's literally the same one — but because of the question. With the same data and the same validation method, predicting price DIRECTION stays glued to chance, and predicting whether VOLATILITY will rise works in 13 out of 13 independent yearly validations, across all four published horizons (3, 7, 14 and 30 days). That fits what market theory says: volatility clusters in streaks and is predictable; direction, in a liquid market, is precisely what competition between participants removes. The practical lesson is uncomfortable but clear: when a model doesn't work, sometimes the problem isn't the model, it's the question.

Build it yourself

The label the model learns is a comparison between two windows. Drag both and check what the model would have had to answer that day.

Label = is future volatility − past volatility > 0?
Volatility difference

In training, future volatility is known because you're looking at the past; in production it isn't, and that's where all the model's work is. The chart below is its answer, day by day, for days it never saw during training.

Illustrative example numbers for practice — not real data.

How it works, step by step

  1. Trailing realized volatility over the last N days (the standard deviation of daily returns) is computed for every day in history.
  2. The label to be learned is a comparison between two windows: will volatility over the next N days be higher than over the last N? It's a yes-or-no question, just like the direction one — only about a different quantity.
  3. The same LightGBM is trained on the same columns: this site's on-chain indicators plus the price-derived technicals. No external data is needed — the project confirmed live that dropping NASDAQ, gold and FX from the original feature set barely moves accuracy.
  4. It's validated year by year, checking it works in each one separately rather than thanks to one particular bull market. That's the 13/13: thirteen independent years, not a thirteen-year prediction horizon.
Honest limitation. It says nothing about direction, and that isn't a quibble: a high probability of more volatility is compatible with an enormous rise or an enormous fall. The project also tried turning it into an exposure rule — cut position when it forecasts a storm — and it didn't work (Sharpe 0.608 against 0.687 for doing nothing), for an economic reason: in Bitcoin the highest-risk periods coincide with the highest-expected-return ones. It has informational value, not allocation value.

Predictions

These are the real predictions stored day by day, each one produced by walk-forward validation: every day's was produced by a model trained only on prior cycles. They're published for transparency, not because there's an edge behind them — for price direction, the project's research found none that is defensible.

P(more volatility) over 30 days

P(more volatility) over 14 days

P(more volatility) over 7 days

P(more volatility) over 3 days

Data last updated on 2026-09-09.

← All models