Time Series Forecasting: Trend, Seasonality and Honest Validation
Time series problems obey different rules from ordinary machine learning. Trend, seasonality, stationarity, baselines and walk-forward validation, explained practically.
Time series forecasting is where ordinary machine-learning habits mislead most. You cannot shuffle the data, you cannot let the future into training, and observations are not independent of one another. Those three constraints reshape the entire workflow.
What a series is made of
Most series decompose into three parts: a long-run trend, a repeating seasonality, and an unexplained residual. That decomposition is the first thing to do before modelling, because seeing which component dominates determines the method.
A series can carry more than one seasonality at once: daily, weekly and yearly patterns may all be present. Retail sales usually have all three.
Stationarity
Most classical statistical methods assume the statistics of the series do not change over time. A trend, or shifting volatility, breaks that assumption. Differencing and log transforms are the usual corrections.
Take baselines seriously
Two baselines are surprisingly strong in time series:
- Naive: tomorrow equals today.
- Seasonal naive: this Tuesday equals last Tuesday.
If a sophisticated model cannot clearly beat those two, the cost of the complexity is being paid for nothing. In many projects the comparison is never run and the results are read far too optimistically.
Honest validation
Random K-fold cross-validation is invalid here; it feeds the future into training. The correct approach is walk-forward validation: each fold trains only on preceding data, tests on the next window, and the window slides forward.
Match the forecast horizon to real use, too. Predicting one day ahead and thirty days ahead are very different problems and should not be measured with one model.
Which metric
| Metric | Strength | Weakness |
|---|---|---|
| MAE | Interpretable, robust | Does not punish large errors |
| RMSE | Sensitive to large errors | Sensitive to outliers |
| MAPE | Comparable as a percentage | Explodes near zero |
| MASE | Scaled against the naive baseline | Less widely known |
Frequently asked questions
Does deep learning beat classical methods?
With many related series, usually yes. On a single short series, classical methods remain highly competitive.
How should holidays and promotions be handled?
Supply calendar effects as explicit features. Expecting the model to infer them from the data is unrealistic, especially on short series.
Is a point forecast enough?
For most decisions, no. Provide a prediction interval; inventory, capacity and risk decisions depend on the size of the uncertainty.
The most common mistake in time series is methodological rather than technical: the wrong split, a missing baseline, and a horizon that does not match real use. For more, see data leakage and cross-validation.
