What is time series prediction
Time Series Forecasting studies the use of past data arranged in chronological order to predict future values. The difference between it and regular regression is that the samples are not independent - the later data is often strongly correlated with the earlier data, and the order itself is a kind of information.
Classical statistical methods
Before the popularity of deep learning, the mainstream approach was statistical models: AR (autoregression), MA (moving average), ARIMA (differential autoregressive moving average, systematized by Box and Jenkins in the 1970s), and Holt Winters exponential smoothing. They have strong interpretability and stable performance on small data, and are still important baselines to this day.
Machine Learning and Deep Learning Methods
Tree models (such as gradient boosting) can use past observations, weeks, holidays, etc. as features to predict. Deep learning further introduces sequence modeling capabilities: LSTM (Hochreiter and Schmidhuber, 1997) can remember longer histories; However, Transformer based models (such as Informer, Zhou et al., AAAI 2021) use attention mechanisms to directly model long sequence dependencies, which are suitable for longer time spans.
Three unavoidable challenges
Trends and seasonality: Data often contains both long-term trends and cyclical fluctuations, and models need to be able to break them down. Non stationarity: Statistical characteristics vary over time, and the training distribution and prediction distribution may not be consistent. Multi step prediction: The later the prediction, the easier it is for errors to accumulate.
How to measure the quality of predictions
Common indicators include MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and MAPE (Mean Absolute Percentage Error). It should be noted that MAPE will distort when the true value approaches 0.
Typical Applications
Sales and inventory forecasting, power load forecasting, financial market analysis, meteorological forecasting, server traffic and capacity planning, etc.
Reference source
Comprehensive compilation of publicly available academic literature and technical documents, including Box and Jenkins' ARIMA method, Hochreiter and Schmidhuber (1997, LSTM), and Zhou et al. (AAAI 2021, Informer).