Analytical models
How this site compares models that try to predict price direction.
What it tries to predict
Not an exact price — that would be a much stronger and much less defensible promise. Instead, for every indicator on this site, the question is: will price be higher within 7, 30, or 90 days? A direction probability is something you can honestly show a confidence interval for.
The models being compared
Logistic regression
Combines every indicator into a single score and turns it into a probability that price will be higher within N days.
Random Forest
Trains hundreds of different decision trees and averages their vote, so one tree's odd opinion can't dominate the final prediction.
Gradient Boosting
Chains decision trees together, where each new one is trained specifically to correct the errors the previous ones made.
LightGBM
The same idea as Gradient Boosting — trees correcting each other in a chain — but built to train much faster with many data columns at once.
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.
How they're evaluated without cheating
The most common mistake when evaluating time-series models is splitting data into "train" and "test" randomly — in a strongly autocorrelated daily series, that lets nearly identical days fall on either side of the split, making the model look far more accurate than it would in real life.
This site instead uses "walk-forward" validation anchored to Bitcoin's halving cycles (2012-11-28 / 2016-07-09 / 2020-05-11 / 2024-04-20): each model trains only on complete cycles prior to the one it's evaluated on. That way the score reflects how the model behaves against a market regime it has never seen, not just a different chunk of time from the same regime.
This is NOT how to evaluate it: if days are split randomly between training and test, near-identical days from the same cycle end up on both sides — the model looks far more accurate than it would in real life.
An honest result worth mentioning
Part of the recent literature reviewed (not exhaustive) found that, for predicting next-day price direction, adding on-chain indicators barely improves on using price alone (open, high, low) — the best model in that study didn't even use on-chain data. That doesn't mean the indicators on this site are useless: every one here is presented as a cycle signal, not a "tomorrow" signal — but it's why no page on this site promises more than its own evidence supports.