ARIMA Models Use Differencing to Forecast Nonstationary Time Series

Summary

The article explains ARIMA models as an extension of ARMA for series with stochastic trends. An integrated series becomes stationary after differencing it a specified number of times; the differenced series is then modeled with autoregressive and moving average terms. The discussion distinguishes this treatment of trends from seasonal patterns and changing volatility, which may require seasonal ARIMA or ARCH/GARCH models.

The examples simulate an ARIMA(1,1,1) process and recover coefficients close to the generating values, then use residual correlograms and Ljung–Box tests to check for remaining serial correlation. For Amazon and S&P 500 daily log returns, the article selects orders by minimizing AIC, checks residuals, and displays multi-day forecasts. It cautions that a good residual fit does not settle broader trading performance: the equity sample begins in 2013 and excludes an earlier high-volatility period, which changes apparent stationarity. Forecasts are shown, but no completed strategy or risk-adjusted performance evidence is provided.

Key ideas

• ARIMA models difference nonstationary series before fitting an ARMA structure to the transformed data.
• The differencing order represents how many transformations are applied to address stochastic trends.
• AIC can be used to compare candidate model orders, while residual diagnostics assess remaining serial correlation.
• The simulated example checks parameter recovery and residual behavior against a known data-generating process.
• Financial time series results depend on the sample period, especially when volatility regimes differ.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.