Forecast intervals for autoregressive models
A point forecast without an interval is incomplete. For an autoregressive model the interval has a closed form and converges to a known width, but it is too narrow when the model is estimated from little data. A simulation compares nominal and realised coverage.
A forecast such as "revenue next quarter: €4.2 million" gives a precise number but no interval, and therefore no indication of its reliability.
Every forecast is a distribution. The single line drawn into the future is its centre, and a complete picture is a fan: narrow where the near future is constrained by the present, wider as the future gets further away. The Bank of England popularised this picture for inflation in 1996, and the format has spread because it is harder to misread than a line. The cone that weather services draw around the forecast track of a hurricane follows the same logic: narrow for the next day, several hundred kilometres wide five days ahead. Fans for interest rates are drawn the same way, from models such as the Vasicek model.
This note describes the fan for the simplest time-series model, why it stops widening, and why the textbook fan is too narrow in practice.
Box and Jenkins (1970) gave the forecast intervals for ARIMA models that every software package now computes. Britton, Fisher and Whitley (1998) explained the Bank of England's fan chart, with bands at successive probability levels shaded from dark to light. Chatfield (1993) reviewed interval forecasts and warned that they are routinely too narrow, chiefly because they ignore the uncertainty in the estimated parameters and in the choice of model. Hyndman and Athanasopoulos (2021) is a freely available modern textbook built around probabilistic forecasting. The model used here is the AR(1) of the autoregressive model of a time series.
The AR(1) forecast
Let with and innovation variance . Standing at time and looking steps ahead, the forecast is
a geometric return from today's value to the long-run mean. Its error is the sum of the shocks still to come, each discounted by the persistence:
Limiting interval width
As grows, the variance tends to , the unconditional variance of the series. Far enough ahead, the forecast no longer depends on the current value and reduces to the unconditional distribution. The more persistent the series (the closer is to 1), the longer the fan keeps opening and the wider it ends. For a random walk () it never stops: the width grows like without limit.
Effect of estimation error
The formula assumes that , and are known. In practice they are estimated, usually from the same short history used for the forecast. Plugging estimates into the formula ignores their uncertainty, and the effect is largest where it matters most, at long horizons, where small errors in and compound. In simulations with and a promised 95% interval:
| History used for estimation | Coverage at | at | at |
|---|---|---|---|
| 20 observations | 90% | 76% | 74% |
| 60 observations | 94% | 88% | 87% |
| 400 observations | 95% | 95% | 95% |
With 20 observations, a "95%" interval 30 steps ahead contains the truth about three times in four. And these numbers still assume the model is right.
For a small business this is a practical matter. A firm that fits its sales model to 20 months of data and plans its cash reserve on the 95% band two and a half years ahead relies on an interval that contains the outcome only about three times in four. The shortfall is not visible in the chart, which looks equally precise whatever the length of the history behind it.
The figure below draws the fan from an estimated AR(1) model and compares its nominal coverage with the realised coverage.
Suggested settings:
- Single forecasts. Press "Reveal the future" a few times with "New history" in between. Most of the time the orange path stays inside the light band; occasionally it falls outside, as a 95% interval implies.
- Coverage. Press "Test 1,000 futures" with 60 observations. Coverage is close to 95% at horizon 1 and falls to the high 80s at horizon 30.
- Shorter history. Set the history to 20 and test again: coverage at long horizons drops to roughly three quarters. Set it to 400: the nominal coverage is achieved at every horizon.
- Persistence. Lower to 0.3: the fan opens to its full width in two or three steps and the forecast is the mean almost immediately. Raise it to 0.98: the fan keeps opening across the whole horizon, and estimation error matters even more.
Derivation of the variance. Subtract and iterate the model times: . The first term is known at time and is the forecast's deviation from the mean. The shocks are independent with variance , so the error variance is , a geometric series. With normal shocks the error is normal, which gives the bands with , 1.282 and 1.960 for 50%, 80% and 95%.
A fan chart is only as reliable as its assumptions. Two practices help: stating how much history the model was estimated from, and checking coverage on past forecasts, in the same way that a scan gate's error rates or a risk curve should be checked. Bootstrap and Bayesian methods, like the one in Bayesian autoregression, widen the fan to include parameter uncertainty.
References
- G. E. P. Box and G. M. Jenkins, Time Series Analysis: Forecasting and Control, Holden-Day, 1970.
- C. Chatfield, "Calculating interval forecasts", Journal of Business & Economic Statistics, 11(2), 121-135, 1993.
- E. Britton, P. Fisher and J. Whitley, "The Inflation Report projections: understanding the fan chart", Bank of England Quarterly Bulletin, 38(1), 30-37, 1998.
- R. J. Hyndman and G. Athanasopoulos, Forecasting: Principles and Practice, 3rd ed., OTexts, 2021.