Machine Learning Forecasting: Lag Features, Train/Test Splits & Error Metrics

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Turn a daily demand series into a supervised-learning table with Lag 1, Lag 7 and a leakage-free rolling mean, then measure which lag carries more signal.
• Derive the weekday with WEEKDAY, build 0/1 dummy columns for Friday, Saturday and Sunday, flag holidays from a list and measure the Saturday uplift.
• Split six weeks of demand at a cut-off date, check the test share, compute training statistics without test data and prove there is no overlap.
• Forecast the two test weeks with the naive and the seasonal-naive rule and score them with MAE, RMSE and MAPE, the bar every ML model must clear.
• Fit demand = b + m₁·Lag 1 + m₇·Lag 7 on four training weeks with LINEST, pull each coefficient out with INDEX, read R² and forecast the next morning.
• Apply the trained coefficients to 14 unseen days, compare the model's MAE with the seasonal-naive baseline, express it as forecast skill and chart actual vs predicted.
• Run a 12-step walk-forward test: at each step the trend model is refitted on all weeks so far and forecasts the next week. Then you compare it with a model frozen after week 8.
• Fit the full model with TREND, refit it three times with one feature dropped, compare test MAEs and chart how much each feature really helps.