Forecasting · seasonality

How to calculate a seasonal index for inventory forecasting, and keep growth out of it

A seasonal index is twelve numbers that say how far each month sells above or below a normal month. This post builds one from three years of monthly sales, shows why the simple version goes wrong on a growing SKU, fixes it, and uses it to size a purchase order. The SKU and every number in the worked example are invented so the arithmetic can be checked.

The short version
  • An index of 1.25 for June means June sells 25% more than a normal month. The twelve indices sum to 12.
  • The simple index (month average ÷ overall monthly average) can mistake growth for season on a growing SKU: December looks too strong, January too weak.
  • The fix: divide each month by a centred 12-month moving average, average the ratios by month, and rescale to sum to 12.
  • To forecast, divide recent sales by the index, project that rate forward, and multiply by the index of each month the order will sell into.
  • Clean the history first: stockouts, one-off promotions and holidays that change month all leak into the index.

Every seasonal forecast needs an answer to one question: how much more, or less, does this SKU sell in a given month than in a normal one? The seasonal index is that answer as a number. It is one input to a demand forecast, next to the underlying rate, the lead time and the safety stock. The simple way to calculate it gives a plausible answer that is biased in a predictable direction, so it is worth doing the slightly longer way.

Step 1 — Build the index from monthly sales

Worked example (invented SKU and sales)

An invented insulated drinks bottle sells best in summer and gets a second lift as a December gift. It has three full years of monthly units, and it is growing:

MonthYear 1Year 2Year 33-year average
Jan143176226181.7
Feb129172211170.7
Mar168208271215.7
Apr201258320259.7
May243301385309.7
Jun273353444356.7
Jul293367474378.0
Aug268336424342.7
Sep209269336271.3
Oct204255331263.3
Nov263344429345.3
Dec314393502403.0
Year total2,7083,4324,353

It sold 2,708 units in Year 1 and 4,353 in Year 3: about 61% growth in two years.

The simple index, and where it goes wrong

The simple method takes each month’s three-year average and divides it by the average of all 36 months:

simple index for a month = that month’s average ÷ overall monthly average

The 36 months total 10,493 units, so the overall monthly average is 10,493 ÷ 36 = 291.47. January’s index is 181.7 ÷ 291.47 = 0.62, and December’s is 403.0 ÷ 291.47 = 1.38. The full set is in the comparison table further down.

The problem is the calendar. December comes eleven months after January in every year. On a SKU growing about 2% a month, December sits on a base about 24% higher than that year’s January (1.02 to the power of 11 is 1.24). The simple average books that growth as season: late months come out too high, early months too low.

The fix: divide by a centred 12-month moving average

A 12-month moving average contains every calendar month once, so it largely smooths out a repeating annual season and tracks the SKU’s underlying level. Divide a month’s sales by the moving average centred on that month and what is left is the season. Twelve is even, so the average is centred using 13 months with the two end months at half weight:

centred moving average for month t = (½ × month t−6 + months t−5 to t+5 + ½ × month t+6) ÷ 12

For July of Year 1: half of January Year 1 is 71.5, February to December of Year 1 add up to 2,565, and half of January Year 2 is 88.0. The total is 2,724.5, and 2,724.5 ÷ 12 = 227.0. July Year 1 sold 293, so its ratio is 293 ÷ 227.0 = 1.291.

With six months needed on each side, 36 months of data give moving averages from July of Year 1 to June of Year 3, so every calendar month gets two ratios: January to June from Years 2 and 3, July to December from Years 1 and 2.

MonthYear 1 ratio (units ÷ moving avg)Year 2 ratioYear 3 ratioAverage ratioIndex
Jan176 ÷ 254.7 = 0.691226 ÷ 322.9 = 0.7000.6960.70
Feb172 ÷ 260.6 = 0.660211 ÷ 331.0 = 0.6370.6490.65
Mar208 ÷ 265.9 = 0.782271 ÷ 337.5 = 0.8030.7930.79
Apr258 ÷ 270.5 = 0.954320 ÷ 343.4 = 0.9320.9430.94
May301 ÷ 276.0 = 1.090385 ÷ 350.1 = 1.1001.0951.10
Jun353 ÷ 282.7 = 1.249444 ÷ 358.2 = 1.2401.2441.25
Jul293 ÷ 227.0 = 1.291367 ÷ 288.1 = 1.2741.2821.29
Aug268 ÷ 230.2 = 1.164336 ÷ 291.8 = 1.1521.1581.16
Sep209 ÷ 233.7 = 0.894269 ÷ 296.0 = 0.9090.9020.90
Oct204 ÷ 237.7 = 0.858255 ÷ 301.2 = 0.8460.8520.85
Nov263 ÷ 242.5 = 1.085344 ÷ 307.3 = 1.1191.1021.10
Dec314 ÷ 248.2 = 1.265393 ÷ 314.6 = 1.2491.2571.26

The twelve average ratios add up to 11.972, not 12. Rescale them by 12 ÷ 11.972 = 1.0024 to get the index column. The unrounded indices sum to exactly 12; the column shows 11.99 only because each value is rounded to two decimals.

Side by side

Because the sales are invented, we know the monthly pattern they were generated from before growth and a little noise were added, so we can see which index recovers it:

MonthSimple indexMoving-average indexPattern the data was built from
Jan0.620.700.70
Feb0.590.650.65
Mar0.740.790.80
Apr0.890.940.95
May1.061.101.10
Jun1.221.251.25
Jul1.301.291.30
Aug1.181.161.15
Sep0.930.900.90
Oct0.900.850.85
Nov1.181.101.10
Dec1.381.261.25

The moving-average index is within 0.01 of the pattern in every month. The simple index is 0.08 too low in January and 0.13 too high in December, an error that climbs across the year: the fingerprint of growth. On a shrinking SKU it runs the other way.

On a growing SKU, the simple index reads growth as season. The moving average takes the growth out before the months are compared.

A check before you use it

This index is multiplicative: it assumes the seasonal swing grows with volume. Check that. Here the best-to-worst month gap was 293 − 129 = 164 units in Year 1 and 474 − 211 = 263 in Year 3, 1.60 times larger, while yearly volume grew 4,353 ÷ 2,708 = 1.61 times. If your swing stays the same size in units as volume grows, use an additive index (units above or below normal) instead.

Step 2 — Use the index on the months the order will sell into

The months that matter are the ones a purchase order has to cover, and those come from your lead time. In the example it is the first week of January in Year 4 and the lead time is three months from purchase order to sellable stock. An order placed now is on the shelf in early April and the next one arrives in early July, so this order has to cover April, May and June.

Take the season out of recent sales

Divide each recent month by its index to get the deseasonalised rate, what the SKU would have sold in a month with no season. For October to December of Year 3:

  • October: 331 ÷ 0.85 = 389.4
  • November: 429 ÷ 1.10 = 390.0
  • December: 502 ÷ 1.26 = 398.4

Their average, 392.6 a month (using the two-decimal indices), is the underlying rate, centred on November of Year 3.

Project the rate forward

One simple estimate of the underlying growth comes from the same division over each full year. Deseasonalised, Year 2 adds up to 3,407.5 and Year 3 to 4,320.3, so the underlying rate grew 4,320.3 ÷ 3,407.5 = 1.268 times in a year. The twelfth root of 1.268 is 1.01997, so about 2.0% a month.

Whether that growth continues is your judgement, not something the index can tell you. This example assumes it does.

April is five months after November, May six and June seven:

MonthUnderlying rateIndexForecast units
Apr Year 4392.6 × 1.019975 (5 months of 2.0% growth) = 433.40.94407.4
May Year 4392.6 × 1.019976 (6 months of 2.0% growth) = 442.11.10486.3
Jun Year 4392.6 × 1.019977 (7 months of 2.0% growth) = 450.91.25563.6
Total1,457.3

(Growth is applied unrounded: 1.01997 a month.)

What the simple index would have ordered

Run the same steps with the simple index and the order comes out at 1,305.7 units, about 152 fewer. The error compounds. The simple index is too high for October to December, so dividing by it understates the underlying rate (365.0 instead of 392.6); then it is too low for April to June, so multiplying cuts the forecast again. Had the order covered October to December, the two methods would differ by about 4 units, because the inflated late-year indices sit on both sides and cancel. That is how the simple index can look fine on one order and be 10% short on the next.

The 1,457 units are expected demand. The order also needs safety stock from your own method and service target. A peak can raise the uncertainty the buffer protects against, but do not scale safety stock by the index; recalculate it with your method’s demand and lead-time inputs. Subtract the stock you expect to have when the order arrives, and that is the order quantity.

Step 3 — Four things that break the index

Months when the SKU was out of stock

A stockout month measures your stock, not your demand. Suppose (invented) an August was out of stock 12 of its 31 days and sold 210 units. Leave it out, or scale it to the days in stock and mark it estimated:

estimated month = units ÷ days in stock × days in month = 210 ÷ 19 × 31 = 342.6

Use that as an estimate only: a short in-stock window can exaggerate demand as easily as a stockout suppresses it. Correct the month before computing the moving averages too: a raw stockout month sits inside thirteen of them, so it drags down its neighbours’ ratios as well as its own.

One-off promotions

A promotion that will not repeat on the same dates is not season. If November of Year 2 carried a promotion you estimate added 60 units (invented), take them out before building the index (344 becomes 284) and plan the promotion separately, as in a promotion forecast. A sale that repeats in the same month every year is part of that month’s season and stays in.

Holidays that change month

Easter can fall in March or April, and Lunar New Year in January or February. If your SKU sells on the holiday, the units move with it. One heuristic: build the index for the two months as a block, then split it by where the holiday falls in the year you are forecasting. In an invented case, March and April have indices of 0.79 and 0.94, a block of 1.73. When the holiday fell in March, March took 52% of the block’s sales; when it fell in April, 42%. Next year it falls in March, so March gets 0.52 × 1.73 = 0.90 and April 1.73 − 0.90 = 0.83. The block total is unchanged, so the year still sums to 12.

A holiday can move supply too: if your supplier closes for Chinese New Year, the reopening date moves your lead time, and with it the months your order has to cover.

A SKU with less than two years of history

With 18 months of sales the moving average exists for only six months, all from one year. That is not enough for its own index. One heuristic is to borrow a pooled index from SKUs you already sell that share the same season driver (the same buyer and reason to buy, not only the same category label) and average their indices month by month. If you adjusted any month by hand along the way, rescale the twelve to sum to 12 again. Check the new SKU’s own months against it as they arrive, and switch to its own index once it has two full years.

What to do this week

  1. Pull 24 to 36 months of monthly units for your seasonal SKUs.
  2. Correct or remove stockout months, one-off promotions and moving-holiday months.
  3. Divide each month by its centred moving average, average by calendar month, rescale to 12.
  4. Compare with the simple index. A gap that climbs across the year is growth leaking in.
  5. From your lead time, list the months your next order sells into and forecast those: underlying rate × growth × index.

For a fourth-quarter order, the timing side is in Q4 starts in August.

Questions people ask

How do you calculate a seasonal index?

Divide each month’s sales by a centred 12-month moving average of the sales around it, average those ratios by calendar month, and rescale the twelve so they add up to 12. The quicker version (month average divided by overall monthly average) can mistake growth or decline for season.

What does a seasonal index of 1.2 mean?

That month sells 20% more than a normal month for that SKU, after growth is taken out. An index of 0.8 means 20% less. The twelve indices add up to 12, so a normal month is 1.0.

Why does my seasonal index make December look too strong?

If the SKU is growing and you used the simple average method, growth is leaking in: December comes eleven months after January, so it sits on a higher base. Divide by a centred moving average instead.

How many years of data do I need for a seasonal index?

Two full years gives every month one moving-average ratio; three gives two and a steadier result. With less, borrow a pooled index from similar SKUs with the same season driver until the SKU has its own history.

How do I use a seasonal index to forecast inventory?

Divide recent sales by their indices to get the underlying rate, project it forward with the growth you expect, and multiply by the index of each month your order will sell into, which your lead time decides. Then add safety stock and subtract the stock you expect to have when the order arrives.

Should I include stockout months and promotions in a seasonal index?

Not as they are. Leave a stockout month out, or scale it to the days in stock and mark it estimated. Remove a one-off promotion’s estimated lift and forecast it separately. A sale that repeats in the same month every year is part of the season and stays.

The SKU, all 36 months of sales, the lead time, the stockout, the promotion and the holiday split in this article are an invented worked example, used so the arithmetic can be checked end to end. The pooled index and the holiday block split are heuristics, and whether growth continues is the reader’s judgement. See SKU Compass plans and pricing, or book a strategy call.

Scroll to Top