Forecast Engine
Predictive model suite for crypto markets. Data pre-processing with Polars for high-performance columnar processing, feature engineering with Numba JIT-compiled functions, Prophet models for trend/seasonality/residual decomposition. Rigorous cross-validation, calibrated confidence intervals, performance metrics and interactive Plotly visualizations. Configurable timeframes (daily, weekly, monthly) with forecast horizon up to 20 days. For quantitative research purposes only.
The problem
Forecasting a crypto price is an exercise where the common mistake isn't getting the direction wrong — it's claiming a precision you don't have. A single number 20 days out is useless at best; what's needed is an honest interval, calibrated on how the model actually erred in the past.
How we tackled it
The suite decomposes the series into trend, seasonality and residuals with Prophet, validates with rolling-window cross-validation, and produces confidence intervals calibrated on observed error rather than assumed up front. Configurable timeframes (daily, weekly, monthly) with a horizon up to 20 days. It stays a quantitative research tool: it measures uncertainty, it doesn't promise certainty.
Technical choices
Polars for columnar pre-processing and Numba for JIT-compiled feature engineering, so the whole validation cycle stays repeatable in minutes; scikit-learn for metrics, Plotly for interactive charts, yfinance for data. One rule governs the code: no future value may enter the computation of a past one.
What it includes
Pre-processing and feature engineering pipeline, Prophet models with trend/seasonality/residual decomposition, rolling-window cross-validation, calibrated confidence intervals, error-metric reporting and interactive Plotly charts. Daily, weekly and monthly timeframes with a configurable horizon up to 20 days, across crypto and traditional assets.