Time series and forecasting
Forecasting is predicting a future that has not happened yet from a past that happened once. Stationarity, autoregression and exponential smoothing, honest walk-forward evaluation, and what the big forecasting competitions found when deep learning met simple baselines.
55 min
After this lesson you can
- Define weak stationarity and the autocorrelation function, and recognise a random walk from its ACF
- Fit an AR(p) model by least squares and derive its h-step forecast intervals from ψ-weights
- Write simple exponential smoothing as an exponentially weighted average and say when it degenerates to the naive forecast
- Evaluate forecasts with walk-forward validation, and compute sMAPE, MASE and OWA exactly as the M4 competition did
- Summarise accurately what M4 and M5 found about statistical, machine-learning and hybrid methods
In this lesson
- The data: one past, one future
- Classical models: autoregression and smoothing
- Evaluating forecasts honestly
- Deep learning for time series
- The strong-baseline lesson: what the competitions found
- Open problems
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.