All work

Horizon

A forecasting and statistical learning pipeline

Role
Solo
Team
Solo
Timeline
2026
Status
Repo coming
Stack
Python, Polars, statsmodels
Links
The repo will be public soon.
observed forecast
Schematic, not a real series: a forecast with 80% and 95% prediction intervals.

Problem

Horizon puts forecasting and statistical learning behind one interface, from raw data to validated results.

Context and constraints

My role

Horizon is a solo project. I designed it and wrote all of it.

Approach

  1. Raw data A series goes in as it comes.
  2. Forecasting Naive, drift, ETS, and ARIMA with automatic order selection.
  3. Statistical learning Lasso, smoothing splines and GAMs.
  4. Validation Residual diagnostics, including the Ljung-Box test.
One interface covers every stage.

Horizon is written in Python with Polars and statsmodels. Naive and drift forecasts are the baselines that ETS and ARIMA have to beat.

Results

What I’d do differently

Building Horizon surfaced four bugs that bias results without raising any error, and I fixed each one:

  • Double differencing.
  • Leakage in cross-validation.
  • Mis-scaled regularization.
  • GAM identifiability.

Next steps

Publish the repository.