Time Series Exercises (Practice Pack)

These exercises are designed to make you “industry-ready” at time series forecasting: correct splits, strong baselines, honest evaluation, and clear reporting.

How to use

• Start with Exercise 1–4 (fundamentals + baselines)
• Then do 5–8 (feature engineering + model comparison)
• Finally 9–10 (production thinking)

Use a public dataset (download yourself). Do not commit datasets or model artifacts.

Exercise 1: Plot and diagnose a series

• Plot raw series
• Plot rolling mean/std
• Decompose (trend/seasonality/residual)
• Identify stationarity (ADF test)

Deliverable: a short markdown section with plots + your conclusions.

Exercise 2: Build a naive baseline

Implement and evaluate:

• last value (“persistence”)
• seasonal naive (last week / last year same period)

Report: MAE, RMSE, MAPE/SMAPE (if applicable).

Exercise 3: Proper time split

Create:

• train
• validation (for tuning)
• test (final)

Compare random split vs time-based split and explain why random is wrong.

Exercise 4: ARIMA/SARIMA baseline

• Grid small parameter ranges
• Pick best on validation
• Evaluate on test

Deliverable: metrics table + a plot of predictions vs actuals.

Exercise 5: Feature engineering (lags + rolling stats)

Create features:

• lags: 1, 7, 14 (or dataset-dependent)
• rolling mean/std: 7, 14
• calendar features: day-of-week, month

Train a tree model (e.g., LightGBM/XGBoost if available, else RandomForest) and compare to ARIMA.

Exercise 6: Forecast horizon experiment

Evaluate performance at:

• 1-step ahead
• 7-step ahead
• 30-step ahead

Discuss which horizon is hardest and why.

Exercise 7: Walk-forward validation

Implement walk-forward validation and compare stability vs a single holdout split.

Exercise 8: Deep learning baseline (optional)

• Build an LSTM/GRU baseline
• Compare to your best classical baseline

Focus on leakage prevention and honest evaluation.

Exercise 9: Drift + monitoring plan (production thinking)

Write a short monitoring plan:

• what to log (latency, error rates, input stats)
• drift checks (mean/quantiles, missingness)
• when to retrain

Exercise 10: Final report (portfolio-ready)

Write a 1–2 page report:

• problem statement
• data description
• baselines
• best model
• evaluation
• limitations + next steps