Prediction Window Size

Old Faithful is predicted +/- 10 minutes, a 20 minute window. For years, Grand Geyser was predicted within a 4-hour interval. You can predict Old Faithful within a 4 hour window, too--it's just not impressive. In fact, you can predict Old Faithful with >99% accuracy in a 2-hour window, days or weeks from now. In any 2 hour window, Old Faithful will erupt. You should strive to make your prediction window as small as possible.

Accuracy

Smaller prediction windows are great but they lead to reduced accuracy. Old Faithful can be predicted within a 5-minute window, but those predictions will be accurate less than half the time. Traditionally, the National Park Service has set their standard at 90% accuracy.

Interval Distributions

Few geysers exhibit an interval distribution that follows a beautiful normal curve. Old Faithful and Castle are bi-modal, but knowing the length of the previous eruption limits the following interval distribution to be roughly normal. Riverside is bi-modal. Daisy is skewed-right. Grand can be skewed-right, normal, or skewed-left.

The Trade-off Between Window Size and Accuracy

Old Faithful is predicted +/- 10 minutes with 90% accuracy. I haven't run the numbers, but let's say it could be predicted +/- 8 minutes with 80% accuracy. An interesting exercise perhaps, but not very practical for saving time trying to see the next eruption. But what about Grand Geyser and its +/- 2 hour window with 90% accuracy? At times, it could be reduced to +/- 1 hour (a 50% reduction) with 80% accuracy (only ~10% reduction). Now that brings some valuable information to the geyser gazer!

Stability, Interval "Stickiness", and Prediction CYA

Old Faithful is remarkably stable. It has changed only very slowly since recorded history with no known dormant periods or any other periods of wild behavior. Riverside is also very stable. Castle is mostly stable. Great Fountain is stable but can shift into "wild phase." Daisy seems to "stick" within a tight interval range for a day or two and then change. Grand can do a "jump shift" in its intervals without warning, maintain the new interval range for a week or two or longer and then return or just stay within the new interval range. All predictors have been burned by sudden changes in geyser behavior. This can lead to window expanding to "CYA"--adding in a fudge factor so even if the geyser goes a little haywire, you might still be right. The NPS sometimes sets prediction windows artificially early (especially for Grand Geyser) under the philosophy that it is better to entice people to the geyser too early so they will wait and see it rather than have them show up too late. The statistician in me despises this method, but the NPS's goals are not limited to number accuracy.