Eleven chapters · four parts.
From the principles that decide whether a series can be forecast at all, through ARIMA and exponential smoothing, into modern machine-learning, deep-learning, transformer, and foundation-model approaches.
Foundations
The Evolution and Principles of Time Series Forecasting
A historical and conceptual entry point — from Babylonian and Mesopotamian omen practices through the Mayan Venus tables to Yule, Slutsky, Wold, and the Box-Jen…
- ▸Place modern forecasting methods in their historical context
- ▸Trace the lineage from Yule, Slutsky, and Wold to ARIMA and beyond
- ▸Recognise the foundational assumptions every linear forecasting model still relies on
The Forecastability of Time Series
Before you choose a model, you have to know whether the series can be forecast at all. This chapter defines forecastability, draws on Hyndman, Knight, and Galbr…
- ▸Define forecastability and explain why it bounds achievable accuracy
- ▸Apply entropy-based measures (approximate, permutation, sample) to assess predictability
- ▸Use Lyapunov exponents to estimate forecast-horizon limits in chaotic systems
Forecasting Metrics
The deepest treatment of forecasting metrics in any modern Python book. Point accuracy, probabilistic scoring (CRPS, quantile loss, Winkler), shape metrics, hie…
- ▸Choose between point, probabilistic, and shape metrics for a given task
- ▸Compute and interpret CRPS, quantile loss, and the Winkler interval score
- ▸Apply Forecast Value Added (FVA) to justify forecasting investments
Classical Models
ARIMA Modeling
ARIMA, end-to-end. Box-Jenkins methodology, AutoARIMA implementations across Python, R, and Julia, comparative performance studies, intuition for the unit-circl…
- ▸Apply the Box-Jenkins methodology to a real series
- ▸Choose between AutoARIMA implementations (statsmodels, statsforecast, pmdarima, R, Julia)
- ▸Diagnose ARIMA residuals using ACF, PACF, and the unit-circle picture
Forecasting with Exponential Smoothing
ETS / state-space exponential smoothing, including the often-misunderstood relationship to ARIMA. Mechanics of SES, Holt's, and Holt-Winters, library comparison…
- ▸Choose between SES, Holt's, and Holt-Winters based on data characteristics
- ▸Read and interpret the ETS state-space framework
- ▸Compare ETS implementations across Python forecasting libraries
Evaluation & Features
State-of-the-Art Performance Evaluation
Forecast stability, drift detection, rolling-origin (walk-forward) cross-validation, blocked CV, and the techniques that make CV honest on dependent data — purg…
- ▸Implement rolling-origin cross-validation correctly
- ▸Apply purging and embargoing to prevent leakage in dependent data
- ▸Detect concept drift in production forecasts
Feature Engineering for Time Series
Twelve sections of feature engineering — statistical, autocorrelation, spectral, shape-based, decomposition-based, embedding, multivariate, holiday, robust, and…
- ▸Build statistical and distributional features from a series
- ▸Decide when frequency-domain features outperform autocorrelation features
- ▸Apply TS2Vec / TS-TCC representation learning to short series
Modern Models
Advanced Machine Learning for Forecasting
Where ML beats classical methods and where it doesn't. Limitations of classical methods, advanced feature engineering, trend and seasonality detection, clusteri…
- ▸Frame a forecasting problem so a gradient-boosted tree can solve it
- ▸Choose between local and global ML approaches
- ▸Apply clustering to time series forecasting
Deep Learning for Time Series
Modern deep architectures for forecasting — RNN/LSTM/GRU, the MLP family (N-BEATS, N-HiTS, DLinear, TSMixer), hybrid and modular designs, and emerging non-atten…
- ▸Choose between MLP-based, RNN-based, and attention-based architectures
- ▸Train N-BEATS / N-HiTS / DLinear / TSMixer end-to-end
- ▸Recognise when a simple MLP beats a transformer on real benchmarks
Transformers for Time Series
A clear-eyed survey of transformers for forecasting — what works, what doesn't, and how to read the literature critically. Informer, Autoformer, FEDformer, Patc…
- ▸Place each transformer variant on the Informer → Autoformer → FEDformer → PatchTST → iTransformer → Crossformer → TimeXer map
- ▸Read transformer-forecasting papers critically
- ▸Recognise why simpler baselines often outperform transformers on standard benchmarks
Foundation Models and LLMs for Forecasting
Where the field is going. Foundation models for time series — Chronos, TimesFM, Moirai, TimeGPT, YingLong, Uni2TS — and a deep look at what the GIFT-Eval leader…
- ▸Define what makes a model a foundation model for time series
- ▸Use Chronos, TimesFM, and Moirai for zero-shot forecasting
- ▸Read the GIFT-Eval leaderboard correctly