Time series and forecasting with AI

In short: A time series is a sequence of measurements over time (temperature, share price, power consumption). Forecasting models learn from the past and its rhythm how things continue.

In more detail: What matters is that order counts. You must not mix training and test data but split by time.

In detail

• Components: trend, seasonality (day, week, year), noise, outliers.
• Classic: moving averages, exponential smoothing, ARIMA, Prophet.
• Machine learning: gradient boosting with “lag” features (value of yesterday, a week ago).
• Neural: LSTM/GRU, temporal convolution, transformers (Informer, PatchTST).
• Foundation models: large pre-trained time-series models predict new series without training (zero-shot).
• Uncertainty: good forecasts give intervals instead of a single value.
• Evaluation: MAE, RMSE, MAPE, backtest with a rolling window.
• Limits: markets and events break patterns. Share prices are hardly predictable.

See also: Recurrent networks and LSTM, Metrics: accuracy, precision, recall and F1