Time Series Forecasting

When working with Time Series data, the main goal is often predicting future values based on historical patterns. Can be done:

Requires (seasonality/trends can be handled separately):

Stationary Time Series: because the patterns remain consistent over time (Non stationary makes pattern detection difficult).

May use:

Forecasting using Lags

Time Series Statistical Methods

Traditional models that explicitly capture trend, seasonality, and autocorrelation in time series data.

Classical Methods: Usually Beats SoTA methods

Baseline Forecast
Moving Average Forecast
Exponential Smoothing
ARIMA – Autoregressive Integrated Moving Average.
SARIMA – Seasonal ARIMA for multiple seasonal patterns.
Prophet – Handles seasonality and holidays well.
Multiple Linear Regression
Regression Metrics
Mean Absolute Percentage Error

Implementations:

Forecasting_Baseline.py – Naive or simple baseline forecasting.
Forecasting_AutoArima.py – Automated ARIMA model selection.
See: Forecasting_Exponential_Smoothing.py

Time Series Machine Learning Methods

Modern approaches that use feature-based forecasting or global models across multiple series. These methods require Feature Engineering for Time Series, such as lag features, rolling windows, and time-based features (hour, day, week). Possible Data Leakage if not setup correctly.

ML/State of the Art Methods: