The plain-language guide
What does a forecast range mean, and how do you order from it?
A forecast of 40 kg with a range of 34 to 46 kg tells you where most days will land. It also tells you how often a day will land outside it. Here is how to read the range and how to choose how much to order.
The short answer
A range with 80% coverage means 8 days in 10 land inside it. About 1 day in 10 lands above the top, and 1 day in 10 lands below the bottom. To choose your order, ask which mistake costs you more. If running out costs more, order near the top. If leftover stock costs more, order near the bottom. If the two cost the same, order the middle.
How to read the three numbers
- The middle is the most likely day. In the example it is 40 kg.
- Check the coverage. An 80% range holds 8 days in 10. A 95% range holds 95 days in 100 and is wider. A range printed without its coverage is missing a number, so ask for it.
- Expect some misses. An 80% range still has about 5 Saturdays a year above the top. That is normal. It is what 80% means.
- Pick your order from what each mistake costs. Work out the cost ratio for the item: the margin you lose on each unit you run short, divided by that margin plus the cost of one unit left over. The ratio is the share of days your order should cover.
A worked example
The forecast for Saturday is 40 kg. The 80% range is 34 to 46 kg.
The first item costs ₱60 a kg and sells for ₱80. Each kilo you run short loses ₱20 of margin. Each kilo left over costs ₱60. The cost ratio is 20 ÷ (20 + 60) = 0.25. Order 36.8 kg, below the middle. You will sell out on some Saturdays, and that costs less than throwing stock away.
The second item also costs ₱60 a kg but sells for ₱200. Each kilo short loses ₱140. The cost ratio is 140 ÷ (140 + 60) = 0.70. Order 42.5 kg, above the middle.
When the margin and the cost of a leftover unit are the same, for example ₱40 each, the ratio is 0.50. Order the middle, 40 kg.
Order from a range
36.8
units to order
Check the range against your own weeks
Keep a tally on paper. Write one line per Saturday and mark the days that landed outside the range. After 10 weeks, 2 marks is what an 80% range expects. 1 or 3 marks is also normal. Ten weeks is too short to judge. After 52 weeks, 8 to 13 marks fits an 80% range. 20 marks means the range is too narrow.
Do not pay for the same risk twice
The range covers a busy day. The safety-stock guide covers a late truck. Say you order the top of the range, 46 kg, and then add 3 days of safety stock at 38 kg a day. You now hold 160 kg for a day that needs about 40 kg. Pick the point in the range from the cost ratio. Size safety stock for late deliveries only.
Why some ranges are wider
A forecast for 3 weeks ahead has a wider range than one for next week. A new item with only 3 Saturdays of sales also has a wide range, because one unusual Saturday changes its average a lot. The range gets narrower as more weeks of sales come in.
For the curious: the statistics behind the range
What the range promises
A prediction interval is a range inside which the next observation is expected to land with a stated probability. Hyndman and Athanasopoulos, in chapter 5.5 of Forecasting: Principles and Practice, usually calculate 80% and 95% intervals. For a normal forecast distribution the interval is the forecast plus and minus a multiplier times the standard deviation of the forecast error. The multiplier depends on the coverage.
| Coverage | Multiplier | Range for 38 kg, spread 5 kg | Width |
|---|---|---|---|
| 50% | 0.67 | 34.7 to 41.4 kg | 6.7 kg |
| 80% | 1.28 | 31.6 to 44.4 kg | 12.8 kg |
| 95% | 1.96 | 28.2 to 47.8 kg | 19.6 kg |
The multipliers are Table 5.1 of the textbook. The 38 kg forecast and the 5 kg spread are example figures. The same forecast gives 3 different ranges, so a range printed without its coverage is incomplete. A 95% range is 2.9 times as wide as a 50% range (19.6 against 6.7).
How often a day lands outside
Coverage fixes the miss rate. A range with coverage c misses on 1 - c of days, split in equal halves above and below. Over 52 Saturdays in a year the table shows what to expect.
| Coverage | Days outside, per 100 | Saturdays above the top, per year | Saturdays below the bottom, per year |
|---|---|---|---|
| 50% | 50 | 13 | 13 |
| 80% | 20 | 5.2 | 5.2 |
| 95% | 5 | 1.3 | 1.3 |
An 80% range still has 5 Saturdays a year above its top. A stall that orders to the top of an 80% range will still run short on about 5 Saturdays a year, and the range was built to allow that.
A ten-week check
The check below uses 10 Saturdays against a forecast of 40 kg with an 80% range of 34 to 46 kg. The actuals are invented.
| Week | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| Actual (kg) | 41 | 36 | 47 | 38 | 33 | 40 | 44 | 39 | 35 | 42 |
| Inside 34 to 46 | Yes | Yes | No, above | Yes | No, below | Yes | Yes | Yes | Yes | Yes |
Two weeks in ten fell outside, which is the 20% an 80% range predicts. Ten weeks is a short test. If the true coverage is 80%, the chance of 4 or more misses in 10 weeks is 12.1%, and the chance of 0 misses is 10.7%. Judge a range over 40 weeks or more, and look for a miss rate far from 20%. A single miss is within what 80% allows.
The table shows how many misses to expect as the weeks accumulate. The typical spread is one standard deviation of a count of misses at a 20% miss rate, so a correct 80% range lands within it about 2 times in 3.
| Weeks tracked | Misses expected | Typical spread | Plausible band |
|---|---|---|---|
| 10 | 2.0 | 1.3 | 1 to 3 |
| 20 | 4.0 | 1.8 | 2 to 6 |
| 40 | 8.0 | 2.5 | 5 to 11 |
| 52 | 10.4 | 2.9 | 8 to 13 |
The band is the expected count plus and minus one spread, rounded to the nearest whole week. Keep a tally of weeks and misses on paper. After 52 weeks, 8 to 13 misses is consistent with an 80% range, and 20 would not be.
Where each cost ratio sits
The overordering guide gives the cost ratio: the margin lost per unit short, divided by that margin plus the cost of a unit left over. The ratio is the share of days the order should cover. The top of an 80% range covers 90% of days, since 10% sit above it. The top of a 50% range covers 75%. These are the same multipliers as the table above, and they let you read off a quantity for any ratio.
| Cost ratio | Multiplier | Where it sits | Order for 34 to 46 kg (80%) |
|---|---|---|---|
| 0.10 | -1.28 | Bottom of the 80% range | 34.0 kg |
| 0.25 | -0.67 | Bottom of the 50% range | 36.8 kg |
| 0.50 | 0 | The centre | 40.0 kg |
| 0.70 | +0.52 | Between the centre and the top of the 50% range | 42.5 kg |
| 0.75 | +0.67 | Top of the 50% range | 43.2 kg |
| 0.90 | +1.28 | Top of the 80% range | 46.0 kg |
Why ranges widen
The textbook notes that prediction intervals usually lengthen as the forecast horizon grows. A forecast for next Saturday is tighter than one for 3 Saturdays ahead. As an example, if the spread is 5 kg for tomorrow, 6 kg for 2 days ahead and 7 kg for 3 days ahead, the 80% half-width grows from 6.4 kg to 7.7 kg and then 9.0 kg. The same chapter warns that intervals built on a normal assumption can be inaccurate, and it says so about one of its own examples. Read the table as a guide to the order of magnitude.
What ForeFlux does
ForeFlux shows a forecast per item per branch with a range around it. It says "not enough data yet" for a location until it has 21 days with sales. The order sheet names the deadline per supplier from its lead time. Every number can be overridden, and an override sticks. The calculator above asks which coverage your range has, because the multiplier changes the answer: confirm the coverage before you use the quantity.
Sources
- Hyndman, R. J. and Athanasopoulos, G., Forecasting: Principles and Practice (3rd ed.), section 5.5, Prediction intervals. 80% and 95% intervals, the multipliers 0.67, 1.28 and 1.96 in Table 5.1, and intervals that usually lengthen with the horizon.
- The miss probabilities (12.1% for 4 or more misses in 10 weeks, 10.7% for none) are binomial arithmetic at a 20% miss rate. The forecasts, spreads and weekly actuals are invented for the example.
Questions owners ask
Past the basics.
Why is the range so wide on a new item?
A spread estimated from 3 weeks of sales is unreliable, and the range inherits that. One unusual Saturday in 3 moves both the average and the spread. A range narrows as weeks accumulate. ForeFlux starts showing forecasts at 21 days with sales for a location, which is 3 weeks, and a range built on a short history is wide.
Should I order the middle of the range every time?
Order the middle when a short unit and a spare unit cost the same. If a unit costs ₱40 and earns ₱40 of margin, the ratio is 0.50 and the centre is correct. Cheap items with a high margin belong near the top, and costly items with a thin margin belong near the bottom. Work the ratio out per item once and keep it with the item.
Can a day land outside a 95% range?
Yes. A 95% range misses on 5 days in 100. Over a 365-day year that is about 18 days, of which about 9 are above the top. The range is a statement of probability. A single day outside it is a normal event, and 18 such days in a year is what the range predicts.
See the next order before you need it
Know what you need, before you need it.
The demo needs no signup and no card. It is the real product running on demo data.
Join the waitlist Open the order list