9 min read

Inventory Forecasting Methods: Test Them on Your SKUs

Inventory forecasting methods differ far less overall than demand patterns do, so choose a method per demand pattern, test it on your sales history, and forecast where you order.

By Praveen Nune, Co-Founder & CEO · Updated 2 Nov 2026

A buyer in a warehouse aisle holds a clipboard and looks up at shelves of boxed stock.

Key takeaways

Pick the method per demand pattern, not once for the whole catalogue. In our test, the gap between methods within one pattern was 5 to 22 percentage points. Exponential smoothing won on steady and sporadic items.
Forecast at the level where you place orders. Summing all products cut error from 66% to 21% in our test.
Check bias (the net tendency to forecast too high or too low) every month, not only the error size. Every method in our test over-forecast.
Hold out your latest 12 months and score each method on them. Keep the simplest method that beats "same as last month".

In this article

1Which inventory forecasting methods work best for each kind of product?
2How do you test an inventory forecasting method on your own sales?
3Why does forecasting each product miss by so much, and what fixes it?
4Does a seasonal peak or one-way bias spoil a simple forecast?
5Do you need machine learning to forecast inventory?
6How do we turn a forecast into purchase orders at Arka Inventory?
7Frequently asked questions

On two years of sales largely to wholesale buyers, the gap between the best and worst of seven methods was 7 points of error overall (leaving out same month last year). The gap between smooth and lumpy items was 28 to 51 points, depending on the method.

We found it by testing the common methods on real invoices from a UK online gift-ware retailer. Choosing among methods matters less than two earlier decisions: what kind of demand each item has, and how many items you add together before you forecast.

This page gives you the result first, then a test sheet so you can repeat it on your own sales before you commit to a method, a spreadsheet formula or a software purchase.

It is written for owners, purchase managers and finance teams who reorder SKUs (stock-keeping units, one per product variant), whether they work in a spreadsheet today or are choosing software. Every figure from our test comes with its method, so you can rebuild it.

Which inventory forecasting methods work best for each kind of product?

The best method depends on the product's demand pattern, and that gap is bigger than the gap between methods. In our test, exponential smoothing won on steady (smooth) and sporadic (intermittent) products. Last month's sales won on erratic and lumpy products, where no method forecast a single product well.

How we ran it. We used the Online Retail II invoices from a UK online gift-ware retailer that sells largely to wholesale buyers, 1 December 2009 to 30 November 2011. The data runs to 9 December 2011; we left out the partial December 2011 so the 24 months are complete. We kept only stock codes made of five digits plus up to two letters, which drops codes such as POST, D, M and BANK CHARGES. We also removed cancellations and non-positive quantities, then summed units per product per month. We kept the 2,409 of 4,876 product codes that sold in both the first six and last six of the 24 months. We forecast each month from December 2010 to November 2011 using only earlier months, 12 forecasts per product. This is one UK business, so treat it as a worked example and rerun it on your own data.

How we classed products. Using each product's first 12 months, we worked out two numbers:

ADI (average demand interval): months in the period divided by months with any sales.
CV² (squared coefficient of variation): the squared ratio of the standard deviation to the mean, of the quantities in months with sales.

We used a cut-off of 1.32 for ADI, the value from Syntetos, Boylan and Croston (2005), and a CV² cut-off of 0.49 that we chose. That gave four classes:

Class	Rule	Products	Months with zero sales
Smooth	ADI under 1.32, CV² under 0.49	671	6%
Erratic	ADI under 1.32, CV² 0.49 or more	822	13%
Intermittent	ADI 1.32 or more, CV² under 0.49	407	37%
Lumpy	ADI 1.32 or more, CV² 0.49 or more	509	37%

Results. The score is WAPE (weighted absolute percentage error): total absolute error across all product-months divided by total units sold. Lower is better. We chose it over the more common MAPE (mean absolute percentage error) because MAPE is undefined whenever a month's sales are zero.

Method	All	Smooth	Erratic	Intermittent	Lumpy
Last month's sales	66%	55%	85%	64%	83%
3-month average	66%	52%	90%	60%	93%
6-month average	71%	54%	100%	62%	105%
12-month average	70%	54%	99%	60%	102%
Exponential smoothing, alpha 0.2	66%	51%	93%	57%	95%
Exponential smoothing, alpha 0.4	64%	50%	87%	58%	90%
Same month last year	87%	67%	122%	83%	121%
Croston-type (SBA)	71%	52%	105%	64%	102%

Exponential smoothing blends the previous forecast with the latest actual: new forecast = alpha × last month's sales + (1 − alpha) × previous forecast. A higher alpha reacts faster. Croston-type methods are built for items with many zero months, and SBA is the Syntetos-Boylan Approximation, their bias-corrected form.

Excluding "same month last year", the methods ranged from 64% to 71% overall. Within a class they differed by 5 points (smooth), 7 (intermittent), 20 (erratic) and 22 (lumpy).

Class	What we would do
Smooth	Exponential smoothing, alpha 0.2 to 0.4. Forecast each product.
Intermittent	Exponential smoothing or a 3-month average. Plan the total over the supplier lead time rather than month by month.
Erratic and lumpy	Forecast the group (supplier, category) and cover the product-level miss with safety stock.

How do you test an inventory forecasting method on your own sales?

Hold out your latest 12 months, forecast each of them using only earlier months, and score each method on the held-out months. A chronological test set that is at least as long as the horizon you forecast is the standard way to avoid flattering a method with data it has already seen. These are the steps we ran, each with a pass or fail.

A woman at a desk studies a sheet of past sales with its latest section covered by a paper strip. — Hiding the latest months while you forecast is what makes the test honest.

1
Export monthly units per product for 24 months.
Pass: every product has a number (0 when nothing sold) for every month. Fail: blank months or a different month count per product.
2
Strip cancellations, returns and non-product lines.
Pass: only positive quantities remain. Fail: credit notes or freight lines inside the units.
3
Classify with ADI and CV² on the first 12 months, using the table above.
Pass: every product has a class.
4
Build each method as a column.
Put the 24 months of actual units in B2:B25, with months 13 to 24 (the test months) in rows 14 to 25. Use these formulas for the first test month, row 14, and fill down:
```
3-month average (C14):  =AVERAGE(B11:B13)
Last month (E14):       =B13
Exp. smoothing (D):     D4  =AVERAGE(B2:B4)
                        D5  =$G$1*B4+(1-$G$1)*D4   (fill down to D25)
                        with alpha typed in G1
```
Pass: C14 equals the mean of B11:B13 worked out by hand, and no forecast cell uses the month it is predicting. Fail: any forecast that reads its own month's actual.
5
Score per class over the test rows.
```
WAPE = SUMPRODUCT(ABS(C14:C25-B14:B25))/SUM(B14:B25)
Bias = (SUM(C14:C25)-SUM(B14:B25))/SUM(B14:B25)
```
Pass: a column that forecasts 0 in every month scores a WAPE of exactly 100%, and your formulas give that. Fail: they give anything else.
6
Choose.
Pass: the method's WAPE in that class is lower than "last month's sales", and its bias is close to zero (we aim for within about 5%). Fail: it cannot beat last month. Then use last month, because a method that cannot beat the simple benchmarks is not worth running.
7
Rescore every month.
Pass: you log WAPE and bias per class each month and retest a class when either moves away from its test-period value. Fail: no log, or the method chosen once is never rechecked.

Why does forecasting each product miss by so much, and what fixes it?

Single-product months are noisy, and adding products together cancels much of the noise, so forecast at the level where you actually order. In our test, a 3-month average missed individual product-months by 66%, but the same forecasts summed across all 2,409 products missed total monthly units by 21%.

The error also sat unevenly across products. We ranked products by units sold in the first 12 months and scored the 3-month average on the 12 test months:

A man tips small scattered piles of goods into one large crate on a table. — Pooling products smooths out the swings that make single items so hard to forecast.

Group	Products	Share of test-period units	WAPE
Top 20%	481	73%	58%
Next 30%	723	19%	83%
Bottom 50%	1,205	8%	96%

That is the logic of ABC analysis. Spend your attention on the few products that carry most of the volume, and manage the rest by group and by rule.

The M5 forecasting competition, run on 42,840 Walmart sales series, shows the same pattern at much larger scale. By our calculation from Tables 3 and B of its results paper, the winning team's WRMSSE (weighted root mean squared scaled error) was about 43% lower than the best benchmark at the total level (0.199 vs 0.350) and about 3% lower at product-store level (0.884 vs 0.915). Gains are concentrated where series are added together.

Does a seasonal peak or one-way bias spoil a simple forecast?

Yes. A simple average carries a peak into the month after it, and every method in our test over-forecast on balance. Net bias ranged from +3% (Croston-type) to +21% (same month last year), with last month's sales at +4% and the 3-month average at +7%.

The peak showed up clearly. The retailer's December 2010 sales were 430,111 units. The 3-month average forecast 487,946 for December (13% over) and 484,266 for January 2011, when only 260,057 units sold, which is 86% over. The 3-month average's monthly WAPE was 51% to 67% in ten of the 12 months and 120% to 128% in January and February 2011.

"Same month last year" is the obvious fix for seasonality, but with one prior year of history it was our worst method, at 87% overall and +21% bias. We would not use it until you have at least two prior years, and not for items whose peak moves between months from year to year, such as festival-driven demand. Seasonality matters most in aggregate: in M5, a seasonal naive forecast beat a plain naive one by about 72% at the top level but only about 6% at product-store level.

Two practical rules follow:

Compare bias against error every month. A forecast that is wrong in both directions is noise. One that is wrong in one direction is a purchasing habit, and it is the cheapest to fix.
Let the forecast set the average and let safety stock absorb the miss. See safety stock calculation and the reorder point formula for the lead-time arithmetic.

Do you need machine learning to forecast inventory?

Not to begin. Machine learning won the M5 competition, with the winning method 22.4% better than the best benchmark. But those gains were concentrated at aggregate levels, on tens of thousands of series with extra data such as prices and calendar events. At product level the gap was small.

For sporadic items, on the Royal Air Force spare-parts dataset, simple exponential smoothing and moving averages were as accurate as Croston-type methods at 3 and 6 months ahead (Li and colleagues). On the M5 data in the same paper, Croston-type methods were slightly better. Our test sits closer to the first result. M5 also found that methods built for intermittent demand did not beat simpler ones by a significant margin.

Our recommendation: start with the sheet above. Move to dedicated planning software when keeping hundreds of forecast columns current becomes the bottleneck, not before. If you are comparing tools, see inventory planning software in India.

How do we turn a forecast into purchase orders at Arka Inventory?

Order quantity = forecast demand over the supplier lead time + safety stock − stock on hand − orders already placed. Forecast the lead-time total with the method your test selected, then order only that gap. Forecasting and Purchase Automation at Arka Inventory links your sales, stock levels and forecasts, so you can plan purchases and schedule purchase requisitions or purchase orders against it.

Arka Inventory is cloud software for inventory, warehouse, purchase and cost management, built natively on Salesforce. Typical onboarding takes a minimum of two weeks, depending on data readiness and integration requirements. We offer a 15-day free trial with full access to all features and no credit card. As of November 2026, plans start at $199 a month billed yearly (see our pricing).

Turn your tested forecast into purchase orders

Arka Inventory links your sales, stock levels and forecasts so you can plan purchases and schedule requisitions or purchase orders against them.

Contact sales team

Frequently asked questions

A demand forecast estimates how many units customers will want. An inventory forecast turns that estimate into what to buy, by subtracting stock on hand and orders already placed and allowing for supplier lead time. Get the demand forecast right first.

Yes. Put units in column B, use =AVERAGE(B2:B4) for the 3-month average and =0.3B4+0.7C4 for exponential smoothing. Excel is enough until the number of products makes updating it slow.

No, a tested average or exponential smoothing is a sound base for a small or mid-size business.

Add up expected demand over the supplier lead time, forecast that total with a 3-month average or exponential smoothing, and score it with WAPE.

Enough to hold out a test period: we used 12 months of history before the first forecast and 12 months of forecasts.

Sources

1.Online Retail II dataset, UCI Machine Learning Repository
2.The M5 Accuracy competition: Results, findings and conclusions, Makridakis, Spiliotis and Assimakopoulos, International Journal of Forecasting, 2022
3.Forecasting: Principles and Practice, "Evaluating point forecast accuracy", Hyndman and Athanasopoulos
4.Forecasting: Principles and Practice, "Some simple forecasting methods", Hyndman and Athanasopoulos
5.Feature-based intermittent demand forecast combinations: bias, accuracy and inventory implications, Li, Kang, Petropoulos and Li, arXiv
6.On the categorization of demand patterns, Syntetos, Boylan and Croston, Journal of the Operational Research Society, 2005
7.Reframing demand forecasting: a two-fold approach for lumpy and intermittent demand, Rožanec and Mladenić, arXiv
8.Forecasting and Purchase Automation, Arka Inventory, November 2026
9.Pricing, Arka Inventory, November 2026

Keep reading

Cost of Goods Manufactured in India: Build the Schedule

Build a cost of goods manufactured schedule for an Indian factory, with a worked rupee example, AS 2 exclusions and five checks that prove the figure is right.

Inventory Valuation Methods in India: FIFO or Weighted Average

Indian standards allow FIFO, weighted average and specific identification, not LIFO. See the rules, a worked example and tests for your inventory records.

Tally Inventory Management: Run This Stock Health Check First

Wrong stock in Tally often traces to how vouchers are posted. Use six checks on Tally reports, settings and a sample count to see if you need new entry rules or a new tool.