The tsecon Guide to Time Series Econometrics¶

A free, full-length guide to time series — from your first autocorrelation plot to research-grade structural identification — where every concept comes paired with code that runs.

The guide mirrors the structure of tsecon, a high-performance time series econometrics library (Rust core, Python API) built in this repository. Each chapter teaches the ideas on their own merits — the guide stands alone as a course in time series econometrics — and then shows the exact library calls that put them to work. Where a method is still on the library's roadmap, the code is clearly labeled a preview and linked to its module specification.

Every chapter follows the same ladder: The idea (plain-English intuition, no equations) → the methods (why you care, the math, runnable code, the classic mistakes) → The frontier (where research stands today) → Which method when (a decision table) → further reading (the founding papers).

The chapters¶

#	Chapter	One line
1	Thinking in Time Series	What makes time-ordered data different, stationarity, random walks, and the transformations that tame them
2	Exploring and Diagnosing a Series	Reading ACF/PACF, testing for white noise and unit roots, and the confirmatory stationarity workflow
3	Honest Inference with Dependent Data	Why dependence breaks textbook standard errors — HAC, EWC, and the bootstrap family that fixes them
4	Univariate Models: AR to State Space	The model ladder: AR → ARMA → ARIMA → ETS → the Kalman filter → regime switching
5	Forecasting: Practice and Evaluation	Backtesting discipline, benchmarks that are hard to beat, accuracy measures, and formal comparison tests
6	Volatility: GARCH and Risk	Why volatility clusters, the GARCH family, VaR/ES, and realized volatility
7	Systems: VAR, Cointegration, and Factors	Modeling many series at once: VARs, impulse responses, common trends, and common factors
8	Structural Identification	From correlation to cause: Cholesky, long-run and sign restrictions, narrative shocks, and external instruments
9	Local Projections	The modern impulse-response workhorse, its inference pitfalls, and LP-IV multipliers
10	Bayesian Time Series	Priors as shrinkage, the Minnesota BVAR, samplers you can trust, and posterior impulse responses
11	Nowcasting and Mixed Frequencies	Reading the economy in real time: ragged edges, MIDAS, factor-model nowcasts, and news decomposition
12	Machine Learning for Time Series	Leakage-safe validation, shrinkage vs sparsity, trees and boosting, and an honest look at foundation models
13	Nonlinear Dynamics: Regimes, Thresholds, and State-Dependent Responses	When linearity fails: threshold and Markov-switching systems, generalized impulse responses, and state-dependent local projections
14	Panel Time Series	Heterogeneous panels: fixed effects, the mean-group and common-correlated-effects estimators, and panel local projections and VARs
15	The Term Structure of Interest Rates	Fitting and forecasting the yield curve: Nelson-Siegel, Svensson, and the dynamic Nelson-Siegel

Worked, figure-rich examples for many of these methods live in the gallery; the library's full technical plans live in the module specifications.

Learning paths¶

You don't have to read linearly. Four curated routes:

The beginner path — never touched time series before: 1 → 2 → 4 → 5. You'll finish able to diagnose a series, fit and select a univariate model, and evaluate forecasts honestly. Add 3 when a p-value matters to you.

The forecaster's path — you ship predictions: 1 → 2 → 4 → 5 → 12, with 6 if your target's uncertainty matters (finance, risk) and 11 if your data arrive at mixed frequencies.

The macro-structural path — you ask "what does a shock do?": 1 → 2 → 3 → 7 → 8 → 9 → 10 → 13. This is the empirical-macro toolkit: from VARs through identification to local projections and Bayesian estimation, the sequence most PhD courses spread across two semesters — capped by 13, where the linearity assumption everything else shares is finally relaxed.

The risk path — volatility and tails are your job: 1 → 2 → 3 → 6, then 5 for evaluating VaR forecasts like any other forecast.

How code appears in the guide¶

import numpy as np
import tsecon

rng = np.random.default_rng(0)
y = np.cumsum(rng.standard_normal(300))      # a random walk
tsecon.check_stationarity(y)["recommendation"]   # -> "Difference"

Blocks like this run today against the library. Blocks introduced with "Roadmap preview" show the intended API for methods still being built — each links to the module spec that defines it. Everything the library computes is validated against reference implementations (statsmodels, SciPy, NumPy, arch) down to tight numerical tolerances; the same discipline keeps this guide's claims honest.

Contributing and errata¶

The guide is versioned with the library. Corrections and clarity improvements are welcome — a guide that teaches judgment has to earn trust the same way the library does: by being checked.