Probabilistic Forecasting
Report the whole spread of futures with an honest interval, not one lonely number that pretends to be sure.
Why Does This Exist?
A hospital orders flu vaccines for next month. A point model says 1,500 doses. True demand lands at 1,600 and 100 patients wait. The point was close and still failed, because ordering needs the high side, not the middle.
Probabilistic forecasting reports the spread: the median for planning, the 90th percentile for safety stock, the 10th for the low side. Order 1,692 and the math covers about 90 percent of futures. Order 1,500 and it covers about half.
This page covers distributions and intervals from a forecaster. Single lag baselines live in arima. Smoothing baselines live in exponential smoothing. Scoring the result lives in forecast evaluation. Honest backtesting lives in time series cross validation.
Think of It Like This
A hurricane cone on the evening news
The forecaster never draws one thin line for a hurricane path. One line would claim certainty nobody holds. The screen shows a cone: narrow tomorrow, wide in five days, with the centre track inside.
Mean demand is the centre track. Spread is the cone width. The P10 to P90 band is the cone edge. Ordering from the centre alone is sailing into the line instead of the cone.
Where it stops: a storm cone shows where the centre may go, never how hard the wind blows there. A forecast interval shows where demand may land, never why it lands there.
How It Actually Works
Call the mean forecast, meaning the centre. Call the spread, meaning one standard deviation of future demand. Call , , and the 10th, 50th, and 90th percentiles, meaning the levels that 10, 50, and 90 percent of futures fall below. The band from to is the 80 percent prediction interval.
Two routes produce it. Quantile regression trains one model per percentile with the pinball loss instead of squared error, then reads and straight off the two models. A parametric head predicts and for a Normal curve, then converts to percentiles with fixed multipliers.
Worked example: mean 1,500, spread 150
Take and for next month vaccine demand. For a Normal curve the 90th percentile sits spreads above the mean and the 10th sits the same distance below:
Subtract and add:
The median is 1500 itself. The 80 percent interval is with width . Ordering 1692 covers about 90 percent of futures. Ordering the point 1500 covers about half. That fan from one centre to a 384 wide mouth is the diagram: history ending today, one dashed point path at 1500, two mint edges at 1308 and 1692.
Why the cone widens
Each extra step adds fresh uncertainty on top of the last. Near horizons lean on recent sales, far horizons lean on trend plus noise stacked many times. A flat interval at all horizons either overstates near confidence or understates far risk. If width never grows, the spread model missed the stacking.
Code
Rebuild the vaccine interval from mean and spread with plain Python.
mu: float = 1500.0sigma: float = 150.0z90: float = 1.2816
half_width: float = z90 * sigmap10: float = mu - half_widthp90: float = mu + half_width
print(f"{half_width:.2f}") # -> 192.24print(f"{p10:.0f} {mu:.0f} {p90:.0f}") # -> 1308 1500 1692print(f"{p90 - p10:.0f}") # -> 384Half width prints 192.24, the three levels print 1308, 1500 and 1692, and the interval width prints 384, matching the hand arithmetic. Raise sigma to 200 and watch the mouth widen past 500.
Watch Out For
Scoring a distribution with point errors
Symptom: two interval models tie on root mean squared error while one misses its 80 percent band half the time and the other holds near 80 percent. Fix: score the distribution, not the centre. Use pinball loss per quantile or ranked probability score for the whole curve, and check empirical coverage of the band beside any point metric from forecast evaluation.
Intervals that never widen
Symptom: the 80 percent band holds the same width at one day and thirty days, then far horizons breach constantly. Fix: suspect a fixed variance head or a quantile model starved of horizon features. Plot width against horizon and coverage against horizon, and retrain with horizon aware spread before trusting safety stock.
The Quick Version
- A point forecast of 1,500 vaccines covers about half of futures, which understocks the other half.
- Mean 1,500 with spread 150 gives half width 1.2816 times 150, which is 192.24.
- The 10th percentile is 1500 minus 192.24, about 1308, and the 90th is 1500 plus 192.24, about 1692.
- The 80 percent interval from 1308 to 1692 spans 384 doses and ordering its top covers about 90 percent of futures.
- Quantile models learn P10 and P90 directly, parametric heads learn mean plus spread and convert.
- Honest cones widen with horizon because uncertainty stacks step after step.