Time-Series Forecasting

Forecasting: clean a real-world signal, inspect trend and seasonality, then compare classical models against a holdout window.

Speed:ms/stage
Stage: 1 / 5Model: Holt-WintersMAPE: 3.40%

Source Signal

Raw observations with gaps still visible.

source
422158937566923810910
Jan 2018Jan 2021Dec 2023

Datasets

Six years of monthly sales with trend, seasonality, a few gaps, and anomalies.
Source points: 72 | Missing values: 2

Forecast Model

Walkthrough Focus

Look at the raw timeline first.

Time series work starts with the original signal. We want to spot gaps, spikes, drift, and whether the series feels seasonal before we trust any model.

Forecast Metrics

MAE
324.4
RMSE
453.3
MAPE
3.40%

Signal Controls

Smoothing Parameters

Current Notes

Interpolated 2 missing observations with linear interpolation.
Clipped 4 outliers between the 2th and 98th percentiles.
Holt-Winters additive captures level, trend, and a repeating additive seasonal profile.

Understanding Time-Series Forecasting