All work

cryptolab

Most crypto “signals” promise the price direction. cryptolab checks that with rigorous tests and says plainly when there is no edge.

Client: Our own project
Type: Library and command line
Our part: Models, C++ core, tests
Year: 2026

What it computes

Market efficiency tests, volatility with GJR-GARCH and HAR-RV, market regimes with a hidden Markov model, the return distribution one day ahead and a walk-forward check against simple models.

An example on BTC

On 3,000 hourly BTC/USDT bars the probability of a rise over a day came out at 52%, and the tool says: no edge. GARCH forecasts volatility slightly better than simple models, but the Diebold–Mariano test does not find the difference significant.

A fast core

The heaviest computations, the GARCH likelihood and the Hurst exponent, are written in C++ and bound to Python. A dedicated test checks that C++ and Python give the same answer.

A laboratory instrument

A tape of hourly bars runs through a chain of meters: efficiency, volatility, market regime. The verdict needle rests on “no edge” until the walk-forward test proves otherwise.

Verdict needle: rests on “no edge” until the walk-forward test passesFour meters in a chain: efficiency, GARCH volatility, HMM regime, walk-forwardPaper tape of hourly bars, as candlesA bench model of cryptolab, a Python and C++ (pybind11) toolkit that analyses crypto markets with textbook econometrics. A paper tape of hourly bars runs through four meters: an efficiency gauge (a variance ratio near 1 means prices have no memory), a pendulum that rings with GJR-GARCH-t volatility, a selector lit by the HMM regime probabilities, and a walk-forward test: until the Diebold–Mariano statistic is significant the verdict needle rests on “no edge”. The tape here is simulated from parameters measured on the real bars, a volatility persistence of 0.971 and an annualised volatility of 18.7%; it is not the bars themselves. A click plants a synthetic edge in the tape to show the needle can move. On the real BTC/USDT run (3000 hourly bars) it did not: P(up) 0.522, Diebold–Mariano p 0.289 over 215 walk-forward windows, verdict no edge. This is not financial advice.

Screens

BTC/USDT analysis results laid out as a panel from the tool’s output

Under the hood

Python with a C++ core through pybind11.
GJR-GARCH-t, HAR-RV, a hidden Markov model, variance-ratio and Hurst tests.
Checks without look-ahead: walk-forward, Diebold–Mariano, purged cross-validation.
An exchange rate limiter and an offline demo on synthetic data.
25 tests, including a C++ vs numpy check. Not investment advice.

Python · C++ · pybind11 · NumPy

Shall we talk about your project?

Discuss your project