sktime Shows 0.584°C MAE in 7-Day Python Forecasting Workflow
Updated
Updated · KDnuggets · Jun 15
sktime Shows 0.584°C MAE in 7-Day Python Forecasting Workflow
1 articles · Updated · KDnuggets · Jun 15
Summary
A 90-day hourly HVAC sensor example in Python used sktime to forecast the final 168 hours, with an ExponentialSmoothing pipeline posting 0.584°C MAE and 2.40% MAPE.
The workflow split data chronologically, defined an absolute ForecastingHorizon, then chained linear imputation, detrending and 24-hour deseasonalizing in a TransformedTargetForecaster before prediction.
An ARIMA(1,1,1) model swapped into the same preprocessing pipeline with almost identical results—0.586°C MAE and 2.41% MAPE—highlighting sktime's interchangeable estimator design.
Cross-validation with an expanding window produced a mean MAE of 0.606°C, reinforcing that the approach generalized across multiple time windows rather than a single holdout test.
The article positions sktime as a scikit-learn-style library for time-series forecasting, classification, regression and clustering, built around pandas-based series, panel and hierarchical data structures.