Time Series Forecasting: ARIMA and Seasonal Models
Apply ARIMA and seasonal time series models for actuarial forecasting on Exam MAS-I.
Stationarity and Differencing
Time series analysis begins with the concept of stationarity. A stationary time series has a constant mean, constant variance, and autocorrelation that depends only on the lag (time separation), not on the time itself. Many real-world time series are non-stationary: claim frequencies may trend upward, premium volumes may grow, and loss ratios may fluctuate with economic cycles. Differencing transforms a non-stationary series into a stationary one by computing W_t = Y_t - Y_{t-1}.
First-order differencing (d=1) removes a linear trend. Second-order differencing (d=2) removes a quadratic trend. In practice, d=1 is sufficient for most actuarial time series. Over-differencing (using d larger than necessary) introduces unnecessary complexity and can distort the autocorrelation structure. The Augmented Dickey-Fuller (ADF) test formally tests for stationarity by evaluating whether the series has a unit root.
Autoregressive (AR) Models
The AR(p) model expresses the current value as a linear combination of p past values plus a random error: Y_t = c + phi_1 * Y_{t-1} + phi_2 * Y_{t-2} + ... + phi_p * Y_{t-p} + epsilon_t, where epsilon_t is white noise with mean 0 and variance sigma^2. The AR(1) model Y_t = c + phi * Y_{t-1} + epsilon_t is the simplest case. When |phi| < 1, the process is stationary with mean mu = c/(1-phi) and autocorrelation rho_k = phi^k at lag k, which decays exponentially.
The partial autocorrelation function (PACF) is the key diagnostic for AR models. For an AR(p) process, the PACF is nonzero at lags 1 through p and zero for all lags beyond p. This "cutoff" pattern identifies the order p. The ACF of an AR(p) process, by contrast, decays gradually (tails off) without a sharp cutoff.
Moving Average (MA) Models
The MA(q) model expresses the current value as a linear combination of current and past random errors: Y_t = mu + epsilon_t + theta_1 * epsilon_{t-1} + ... + theta_q * epsilon_{t-q}. The MA(1) model Y_t = mu + epsilon_t + theta * epsilon_{t-1} has autocorrelation rho_1 = theta/(1 + theta^2) at lag 1 and rho_k = 0 for all k > 1. MA processes are always stationary regardless of parameter values.
For MA(q) processes, the ACF cuts off at lag q (is zero for lags beyond q), while the PACF tails off. This is the opposite pattern from AR models and provides the diagnostic tool for identifying MA order. An MA(q) process has "memory" of only the last q random shocks.
ARIMA Models
The ARIMA(p,d,q) model combines autoregressive terms, differencing, and moving average terms. After differencing d times to achieve stationarity, the resulting series follows an ARMA(p,q) model. The ARMA(p,q) process has both ACF and PACF tailing off, which distinguishes it from pure AR or pure MA processes. The Box-Jenkins methodology provides a systematic approach to model identification, estimation, and diagnostic checking.
Model identification begins with examining the ACF and PACF of the differenced series. If the PACF cuts off at lag p and the ACF tails off, an AR(p) model is appropriate. If the ACF cuts off at lag q and the PACF tails off, an MA(q) model is appropriate. If both tail off, an ARMA model is needed, and the orders p and q must be determined by trying several combinations and comparing AIC or BIC values.
Seasonal ARIMA (SARIMA)
Seasonal ARIMA extends the framework to capture periodic patterns. The notation ARIMA(p,d,q)(P,D,Q)_s specifies both non-seasonal components (p,d,q) and seasonal components (P,D,Q) with seasonal period s. Seasonal differencing computes W_t = Y_t - Y_{t-s} to remove seasonal patterns. For monthly data with annual seasonality, s=12, so seasonal differencing subtracts the value from 12 months ago.
The seasonal ACF and PACF show patterns at the seasonal lags (s, 2s, 3s, ...). A seasonal AR(1) process shows exponential decay in the ACF at seasonal lags and a PACF cutoff at lag s. A seasonal MA(1) process shows an ACF cutoff at lag s and tailing off in the PACF at seasonal lags. In practice, simple seasonal models like ARIMA(1,1,1)(1,1,1)_12 capture both short-term and seasonal dynamics effectively for many actuarial applications.
Forecasting and Forecast Intervals
ARIMA models produce forecasts by iterating the model equation forward. For an AR(1) model, the one-step-ahead forecast is Y-hat_{T+1} = c + phi * Y_T. Multi-step forecasts are computed recursively: Y-hat_{T+2} = c + phi * Y-hat_{T+1}. As the forecast horizon increases, forecasts converge to the unconditional mean of the series.
Forecast intervals quantify uncertainty. The (1-alpha)% forecast interval is Y-hat_{T+h} +/- z_{alpha/2} * sigma_h, where sigma_h is the forecast standard error at horizon h. For AR(1), sigma_1 = sigma_epsilon and sigma_h increases with h as uncertainty accumulates. Forecast intervals widen with the forecast horizon, reflecting the decreasing precision of longer-range predictions.
Common Pitfalls
A frequent mistake is identifying model orders from ACF/PACF plots of the original (undifferenced) series. Always difference first to achieve stationarity, then examine the ACF and PACF of the differenced series. Another common error is confusing the cutoff and tailing-off patterns: AR models have PACF cutoff and ACF tailing off, while MA models have the opposite. Mixing these up leads to incorrect model specification.
Students also sometimes forget that ARIMA forecasts converge to the mean for long horizons. This means ARIMA models are best suited for short-to-medium-term forecasting. For long-term projections, structural models or judgment-based approaches may be more appropriate.
Study Tips
Exam MAS-I tests model identification from ACF/PACF plots, parameter estimation, forecasting, and forecast interval construction. Practice reading ACF and PACF plots and mapping patterns to model types. Know the theoretical ACF and PACF for AR(1), AR(2), MA(1), MA(2), and ARMA(1,1) processes. Be prepared to compute one-step and multi-step forecasts from fitted models. Understand the Box-Jenkins methodology as a complete workflow from raw data to validated forecasting model.