← AI Terminology

Time Series Forecasting

Time series forecasting is the ML task of predicting future values of a sequentially ordered, time-indexed dataset — using historical patterns (trends, seasonality, cycles) to produce point estimates or probabilistic forecasts for future time steps.

It is one of the most commercially valuable ML applications across finance, supply chain, and operations.
Why It Matters in AI
Accurate forecasting has direct business value: a 1% improvement in demand forecast accuracy can reduce inventory costs by millions for a retailer; energy grid operators need 24-hour electricity demand forecasts to dispatch generation efficiently; hedge funds make billions from time-series-based trading signals. Traditional methods (ARIMA, Exponential Smoothing) were standard for decades; deep learning (LSTMs, Transformers, N-BEATS) and now LLM-based zero-shot forecasters are transforming the field.
Key Points
Aspect Description
Classical ARIMA (AutoRegressive Integrated Moving Average), Exponential Smoothing, Prophet — interpretable, fast
Evaluation MAE, MASE, RMSSE, sMAPE — scale-independent metrics for cross-series comparison
Deep learning LSTM, Temporal Convolutional Network, N-BEATS, PatchTST — learn patterns from many series
Probabilistic Forecast distribution not just point estimate — quantiles, Gaussian NLL, conformal prediction
Foundation models TimesFM (Google), Moirai, Chronos — pre-trained on millions of time series; zero-shot forecasting
Transformer-based Informer, Autoformer, iTransformer — attention-based for long-horizon forecasting
Simple Analogy
A weather forecaster for any data: they study years of historical patterns (seasonality, trends, anomalies), build a model of how the metric behaves, and project it forward. Good forecasters express uncertainty ("70% chance of rain") — probabilistic forecasting is the equivalent for demand, prices, or energy.
Common Usage Examples
  • Prophet: from prophet import Prophet; model = Prophet(); model.fit(df); forecast = model.predict(future)
  • from statsmodels.tsa.arima.model import ARIMA; model = ARIMA(y, order=(1,1,1)).fit()
  • N-BEATS: pip install neuralforecast; NeuralForecast(models=[NBEATS(h=12)], freq="M")
  • Chronos: from chronos import ChronosPipeline; pipeline.predict(context, prediction_length=12) — zero-shot
  • PatchTST: from transformers import PatchTSTForPrediction — transformer-based forecasting model
Summary
In short: Time series forecasting predicts future values from historical temporal patterns — one of the highest-value ML applications, evolving from classical ARIMA to deep learning and now foundation models that forecast zero-shot across any domain.