Time-Series Models for Financial Data › Limitations and Practical Considerations

Time-series models help analyze serial dependence, but they have limitations with financial data.

Linear structure assumption: ARMA models capture only linear dependencies. As we observed in the chapter on stylized facts, financial returns exhibit significant nonlinear features. While squared returns may be highly autocorrelated (indicating predictable volatility patterns), raw returns often show minimal linear autocorrelation. This means ARMA models may suggest returns are nearly unpredictable when substantial predictability exists in higher moments. The next chapter on GARCH models addresses this limitation by modeling the conditional variance process.

Stationarity requirements: The assumption that the data-generating process remains stable over time is often violated in financial markets. Structural breaks occur due to regulatory changes, market crises, and regime shifts. An ARIMA model estimated on pre-2008 data would perform poorly during the financial crisis. Adaptive estimation windows and regime-switching models partially address this, but no mechanical solution exists for structural change.

Gaussian innovations: The standard ARIMA framework assumes normally distributed innovations, yet financial returns consistently exhibit fat tails. This matters most for risk applications: forecast intervals based on normal distributions systematically underestimate tail risk. Using Student-t or other fat-tailed distributions for the innovation process improves this, though at the cost of added complexity.

Limited forecasting horizon: For typical equity returns with autocorrelation coefficients near zero, ARIMA models provide little forecasting value. The optimal forecast quickly converges to the unconditional mean. ARIMA models work better for series with persistent dynamics, such as interest rate spreads, volatility measures (the square of returns), and macroeconomic indicators.

Model selection uncertainty: The Box-Jenkins methodology requires judgment calls about differencing order, lag selection, and handling of outliers. Different practitioners may arrive at different specifications from the same data. Information criteria help but don't eliminate subjectivity. In high-frequency trading applications, this model uncertainty itself becomes a risk factor.

Despite these limitations, ARIMA models established the foundation for modern financial time-series analysis. The concepts of stationarity, autocorrelation structure, and systematic model identification remain central even as we extend to more sophisticated approaches.