Data such as stock prices or the yen-dollar rate are often “non-stationary,” meaning their average and variance shift over time. Robert Engle proposed the ARCH model, which captures the way large swings in returns tend to cluster, thus quantifying risk that varies over time. Clive Granger introduced the idea of cointegration: even if two series wander randomly, a certain combination of them may remain stable, reflecting a long-run equilibrium. This led to error-correction models that link short-run volatility with long-run balance. Financial institutions use these tools to meet Basel risk-capital rules and firms refine price forecasts and inventory control. Scholars also avoid spurious correlations, so empirical economics has advanced dramatically.

Engle (1982) introduced the ARCH(q) process, assuming that the conditional variance of regression errors follows an autoregressive structure. He supplied maximum-likelihood estimation and an LM test for residual heteroskedasticity. Subsequent extensions—GARCH(p,q), EGARCH, TARCH, stochastic volatility models—are now standard in financial risk management and asset pricing. Granger and Engle (1987) defined cointegration as the presence of an I(0) linear combination among I(1) variables, and proposed the ADF residual test with a two-step estimation of the error-correction model. This unified the empirical testing of term-structure theory, purchasing-power parity, and the permanent-income hypothesis. Within the VAR/VECM framework, Johansen’s maximum-likelihood method estimates the cointegration rank for multiple series, enabling dynamic policy analysis and international finance forecasting. Today these techniques are embedded in Python and R libraries, making them core infrastructure for applied economic analysis.

In high-frequency finance, conditional variance exhibits long memory and leverage effects, motivating FIGARCH and APARCH specifications in practice. Bayesian estimation and particle-filter SV-GARCH hybrids improve VaR and Expected Shortfall accuracy. For cointegration, structural-break robust tests such as Gregory–Hansen and non-linear Threshold VARs allow asymmetric adjustment speed to be evaluated. With large panels, Pedroni and Westerlund panel-cointegration tests and factor-augmented VECMs facilitate analysis of international portfolio diversification and global value chains. Machine-learning hybrids feed GARCH residuals into LSTMs to enhance risk forecasts. Realized GARCH and HEAVY models incorporate high-frequency realized measures, mitigating micro-structure noise; thus, the ARCH family remains an active research frontier.