# Demand Forecasting for Retail: Methods, Data and Benefits

> Learn how demand forecasting for retail works, which methods to use, the data it needs, and how to cut stockouts and overstock with better inventory plans.

Source: https://www.electe.net/post/demand-forecasting-for-retail

Site guide: https://www.electe.net/llms.txt

A stockout usually gets blamed on a supplier. Most of the time it starts closer to home. A [retail inventory study presented at the IEOM Society's 2025 Bangladesh conference](https://ieomsociety.org/proceedings/bangladesh2025/372.pdf) indicates that **51% to 73% of supermarket stockouts originate inside the retail operation**, with inaccurate forecasting (including underestimated demand) and ordering errors among the major causes.

That finding reframes the work. **Demand forecasting for retail** isn't a contest to pick the most advanced algorithm. It's an operating discipline: pulling honest demand signals out of data that stockouts and promotions have already distorted, matching the method to how each product sells, and wiring the output into reorder points a buyer will use on Monday morning.

This guide covers the whole loop. You'll see what the forecast has to produce, which inputs carry the weight, how six common methods compare by retail situation, why slow movers need a different architecture, how stockpiling and channel shifts inflate apparent promotion lift, which error metric to trust for which decision, and a five-step check to run before your next buying cycle.

## What Demand Forecasting Means on the Shop Floor

A weather forecast earns trust when it is local and timed: the chance of rain in your city on Tuesday afternoon is enough to decide whether to carry an umbrella. Demand forecasting does the same job for inventory. It estimates how many units of a specific item customers will buy in a specific store during a specific week, so a buyer orders the right quantity rather than a plausible one.

The five steps below are the whole idea in miniature, from the prediction to the purchase order it should trigger.

The target is demand, not the sales tally the till happened to record. A zero logged while the shampoo shelf was empty is censored demand; a promotional spike can be borrowed from the next fortnight or pulled from a nearby store. Category averages and monthly reports hide those item-store problems, so the signal has to stay granular and refresh often.

The working loop looks like this:

1. **Collect** sales and stock history by item, store and day, including the days a shelf was empty.
2. **Clean and align** it: one unit of measure, one calendar, no duplicated rows when a store system restates a week.
3. **Fit and forecast** demand per item-store, at the granularity your ordering decision needs.
4. **Translate the forecast into orders and safety stock**, using lead times and the service level you've promised.
5. **Measure errors and adjust**, item by item, rather than once a year at budget time.

Each step protects against one of three costs. Under-forecast and you lose sales you cannot recover, because the shopper buys the substitute or leaves. Over-forecast and cash sits in stock that has to be marked down. Get the quantity roughly right but the timing wrong and you pay both ways, with freight premiums on the rush order and storage on the surplus.

> **The rule that saves the most money:** a forecast only earns its keep when a replenishment rule acts on it. A perfect number sitting in a report changes nothing on the shelf.

## The Data Inputs and Why Stockouts Distort Them

Before you compare models, look at what goes in. The inputs that move retail accuracy most are predictable and boring, and many teams are missing two or three of them.

- **Sales history by item, store and day.** Anything coarser hides the promotion weeks and weekend spikes that carry the signal.
- **On-hand inventory and stockout duration.** Without them you can't tell a quiet week from an empty shelf.
- **Prices and discount depth.** A 10% cut and a 30% cut don't produce the same response curve, so store the shelf price shoppers pay, not the list price.
- **The promotion calendar**, including display placement and feature advertising.
- **Holidays and local events.** A school fete two streets away moves more bottled water than a national holiday does.
- **Weather**, particularly for fresh, seasonal and garden categories.
- **Competitor pricing**, wherever you can observe it consistently.
- **Supplier lead times and delivery delays**, which decide how much cover you need.

### Why an Empty Shelf Writes a False Zero

Say your shampoo sells 40 units a week on average and the shelf sits empty for three days. Those days record zero sales. A model trained on that history learns that demand falls whenever inventory runs out, which is the opposite of the truth. Shoppers who found nothing didn't change their minds; they went elsewhere.

Supplier delays, slow replenishment and your own ordering choices censor sales in exactly the same way. The model has no way to tell a lost sale from a lost interest, and it will keep under-ordering the item that sells out fastest.

The failure is well documented. A [2026 review of machine learning in retail forecasting](https://link.springer.com/content/pdf/10.1007/s10462-026-11723-2_reference.pdf) identifies incomplete datasets, fragmented data pipelines, changing customer behavior, promotions, competitor reactions and seasonality as causes of biased, brittle models. The same review cites a survey of **661 inventory professionals** in which **51% sometimes proceeded with forecasts they already considered unreliable** because no better alternative existed, while tariff and trade-policy changes (19%) and inflation or commodity-price swings (17%) ranked among the biggest disruptors.

### The Four-Step Fix, in Order

1. **Flag every stockout day** in the sales history, with hours or days the item was unavailable.
2. **Estimate the lost sales** for those days from comparable items in the same store and week.
3. **Tag each period** as available or constrained before any training run starts.
4. **Train on the adjusted history and compare results on held-out weeks**, never on the weeks you patched.

For most SMEs this work moves accuracy further than switching algorithms. A model can only learn what the record contains; if the record says demand collapsed during a stockout, a fancier architecture just learns the wrong thing faster. Our [ELECTE guide to multi source integration](https://www.electe.net/post/multi-source-data-integration) covers the plumbing that gets sales, stock and promotion data into one usable table.

## Forecasting Methods Compared by Retail Situation

There's no best method, only a method that fits the situation. The table below is the comparison I'd want pinned above a planning desk.

MethodBest retail situationData it needsStrengthWatch-outMoving average or exponential smoothingStable, high-volume items1-2 years of weekly salesSimple, cheap and easy to explainMisses promotions and seasonal shiftsSeasonal models such as SARIMAItems with clear weekly or yearly cyclesTwo or more seasons of historyCaptures trend and seasonalityStruggles across thousands of series at onceCroston-family methodsIntermittent, slow-moving SKUsSales history with many zero periodsSeparates how often a sale happens from its sizeWrong choice for steady sellersRegression with promotion and price featuresPromotion-driven categoriesPromotion calendar and price historyShows how each driver moves demandBreaks down when promotions are sparseTree ensembles such as XGBoost or LightGBMLarge item-store portfoliosSales, calendar, promotions and pricingLearns across many related seriesNeeds clean data and ongoing monitoringHybrid LSTM-XGBoost ensemblesMulti-category chains with rich signals36 months of point-of-sale and promotion dataReduced mean absolute percentage error by an average of 28.6% at a four-week horizon in [one study of 14 retail organizations](https://sciety.org/articles/activity/10.31235/osf.io/4z9be_v1)Heavier to build and explain

The pattern behind the rows matters more than any single one. A [peer-reviewed review of retail forecasting research](https://pmc.ncbi.nlm.nih.gov/articles/PMC9534457/) reports that global machine-learning models, including feed-forward and recurrent neural networks, gradient-boosted regression trees and random forests, often outperform separate time-series and regression models when retailers forecast large collections of sales series. The advantage grows with the portfolio: thousands of item-store combinations where individual products have sparse, intermittent, seasonal or promotion-sensitive demand.

### What the Comparison Looks Like in a Real Chain

Picture a regional chain with 40 stores. Staple items with steady weekly sales get exponential smoothing, because a category manager can read the number and argue with it. Seasonal apparel with eight promotions a year gets a gradient-boosted model trained across all stores at once, because one store's history is too thin to learn a discount curve from.

The review is equally clear that the older methods aren't obsolete. Their performance stays competitive on smaller datasets, stable products, and settings where interpretability and limited data matter more than the last point of accuracy.

Restraint pays. A [2026 evaluation of U.S. retail demand](https://econpapers.repec.org/article/alkfcsai0/v_3a5_3ay_3a2026_3ai_3a9_3aid_3a13509.htm) on the M5 dataset, covering **25,940 item-store series** across five rolling 28-day windows, cut mean weighted absolute percentage error from **0.9006** for a seasonal-naive benchmark to **0.7022** with a LightGBM model using historical demand, an improvement of about **22%**. Adding retail and operational variables reached only 0.6979, and lagged macroeconomic indicators reached 0.6978, a gain of roughly **0.02%**.

> **Build the baseline first.** A seasonal-naive or exponential-smoothing number that takes an afternoon to produce tells you whether the machine-learning model earned its complexity. Run both on the same held-out weeks and keep the winner.

Model choice should follow the scale and structure of the portfolio, not whether the method is newer. That single principle will save you more money than any hyperparameter search.

## Slow Movers and Sparse Demand Need Their Own Model

Some products don't have a demand level. They have a demand event. A spare part that sells four times a year, a long-tail e-commerce variant, a store-SKU combination that moves two units a month: standard seasonal models handle these badly, and the reason is structural.

Zeros and positive sales come from two different processes. Whether a sale happens at all is one question. How large it is when it happens is another. Croston-type methods estimate those separately, which is why they hold up on zero-heavy retail data. A [retail forecasting benchmark on intermittent demand](https://arxiv.org/pdf/2602.22685) reports Croston at **WAPE 93.15%** and **MASE 0.9852**, ahead of the naive and SARIMAX alternatives it was tested against. Read that WAPE as a relative result on a series full of zeros, not as a good accuracy score in absolute terms.

The failure mode is specific and expensive. A standard regression or seasonal model sees a run of zeros and concludes demand is low and steady, so it sets safety stock accordingly. The slow mover that sells in unpredictable bursts then runs dry while the dead item holds three months of cover. One error costs a sale, the other costs cash and shelf space.

There's a second lesson in the same territory. A [benchmark on sales data with intermittent demand, missing observations and frequent assortment turnover](https://arxiv.org/html/2506.05941v2) found localized tree-based models performed best among the approaches tested, with XGBoost achieving the lowest reported root mean square error of **4.833**. The takeaway isn't "use XGBoost everywhere". It's that when products turn over quickly, models built per product family can beat one generalized model forced to average across incompatible patterns.

### How to Route a Series

1. **Classify before you model.** Compute the Average Demand Interval and the squared coefficient of variation of non-zero demand for every series. The first is the typical gap in periods between sales; the second measures how much the size of those sales swings from one sale to the next.
2. **Route intermittent series** to Croston-family, hurdle or occurrence-size models.
3. **Send stable, high-volume items** to seasonal or machine-learning models, where they belong.
4. **Recheck the classification quarterly.** Items migrate between categories as they age, get discontinued or catch on, and a routing rule that never gets re-run degrades.

> Keep the classification in a table you can inspect. When a category manager asks why a product is being forecast with an odd method, the answer should be two numbers, not a shrug.

## Promotions, Stockpiling and Channel Shifts

Promotional weeks aren't ordinary weeks, and blending them into a baseline corrupts both. Model baseline demand and promotion effects as separate things, then add them.

A [comparison of retail forecasting approaches](https://inria.hal.science/hal-02460504/document) found that decision-tree ensembles outperformed linear regression, penalized linear regression and regression-tree methods for chain-level daily demand when SKUs had **more than five historical promotions**. Five is the useful threshold in that result. Below it there aren't enough examples for a model to learn how discount depth, promotion timing and product-calendar interactions move demand, so stronger regularization or a simpler benchmark does less damage than a flexible model chasing noise.

Give the model explicit features: promotion status, discount percentage, display placement, holiday proximity, and lagged post-promotion demand. The last of these is often left out, yet it is the feature that captures the dip.

### The Dip Nobody Forecasts

Say a 30% discount on shampoo lifts sales for one week, then sales sit below trend for two weeks. Shoppers stocked up; they didn't buy more over the quarter. Treat the whole three-week window as one unit of analysis. Promoted-period error tells you whether you ordered enough. Post-promotion error tells you whether you over-ordered for the rest of the month. Scored separately, both decisions show up. Scored together, the two errors cancel and the report looks clean while you got both wrong.

[Research on the promotional lifecycle](https://ideas.repec.org/a/wly/jforec/v45y2026i1p293-315.html) notes that prior work often overlooks the post-promotional phase, where demand falls because customers stockpiled during the event, and reports that machine-learning methods can improve forecasts across the entire lifecycle rather than only during the promotion itself.

Channel adds a second layer of illusion. A [study of multichannel promotions at a regional Brazilian supermarket](https://prod.org.br/article/doi/10.14488/1980-5411.20250127) estimated demand displacement of **11.23% from physical stores to online** and **19.34% in the reverse direction**, while coordinated promotions reduced cannibalization by **30.04%**. Loyalty incentives produced the largest measured uplift, at **3.5%**. Forecast each channel independently and you'll double-count demand that changed address.

Then hold the evaluation to four rules:

- **Score promoted and non-promoted periods separately**, so a good week can't mask a bad month.
- **Compare against a base-lift or similar causal baseline**, or you're just measuring how well the model learned volume.
- **Evaluate by SKU and by promotion frequency**, because the five-promotion threshold changes what you can expect.
- **Keep a lagged post-promotion feature alive** so the model can express the dip instead of being surprised by it.

If discounts are eating your margin, our guide to [protect margins during discount campaigns](https://www.electe.net/post/promotion-optimization-in-retail) covers the pricing side of the same problem.

## Measuring Accuracy and Turning Forecasts Into Orders

Pick the metric after you pick the decision. Each measure answers a different question and hides something different.

MetricUse it whenWatch out forWAPE (weighted absolute percentage error)Comparing accuracy across items of different sizesWeights high-volume items most heavilyMASE (mean absolute scaled error)Comparing models on intermittent or zero-heavy seriesNeeds a naive benchmark to read correctlyMAPE (mean absolute percentage error)Simple reporting on steady, high-volume itemsUnstable or undefined when actual demand is zeroBiasSpotting systematic over- or under-forecastingOffsetting errors can hide it unless read by periodService levelChecking whether the forecast protects shelf availabilityDepends on replenishment settings, not the forecast alone

Percentage errors break on slow movers. When actual demand is zero, the denominator is zero and the metric either explodes or gets excluded from the average, and both outcomes flatter the model. Report WAPE or a scaled error alongside bias and service level, and read them by segment rather than as one catalogue-wide number.

### From a Number to a Reorder Point

A forecast becomes an order through arithmetic anyone on the team can check. Expected demand over the lead time, plus safety stock sized to your target service level, gives the reorder point.

Say a store expects 40 units over a five-day supplier lead time and holds a 10-unit buffer. The reorder point is 50 units. When on-hand plus on-order drops to 50, the order goes out. If the buyer orders at 40 because the buffer was never agreed, the shelf empties on the first day the truck runs late.

Two routines keep this honest. Set exception alerts when forecast error on an item exceeds your threshold for two or more consecutive periods, so someone looks at the item instead of the dashboard. Run a monthly bias check by category to catch systematic over- or under-forecasting that individual item errors average away.

> A statistically accurate forecast can still cause shortages if order timing, review frequency or safety stock is miscalibrated. Review the forecast and the replenishment rules together, or you'll keep fixing the wrong one.

Our [ELECTE stock data analytics guide](https://www.electe.net/post/inventory-management-logistics) goes deeper on connecting stock positions to the numbers buyers act on.

## Confirm Your Forecast Plan Before the Next Buying Cycle

Before the next order is placed, confirm that the forecast process fits the products and decisions it needs to support. Five checks make the plan practical:

1. **Triage your assortment.** Identify the items that drive most of your revenue, then classify them as stable, promotion-driven or intermittent. A simpler rule may be adequate for the rest.
2. **Check the sales history.** Pull 12 to 24 months of sales by item, store and day, and flag stockout days. With only 12 months available, use the data but expect wider error.
3. **Test against a baseline.** Set a seasonal baseline and compare richer models on held-out weeks. The M5 result above shows why a more complex model should earn its place.
4. **Set the measurement routine.** Choose a metric for each decision, then keep its threshold and review cadence consistent across cycles.
5. **Review repeated misses.** Re-check reorder points for items with recurring forecast errors before placing the order.

ELECTE runs six forecasting methods on user data, scores them against each other and selects the most accurate one. Its free reorder point calculator helps calculate an individual item's reorder point.

Forecasts reduce risk, but they cannot account for every change. A model may miss a new product launch or a promotion with no comparable sales history. Keep a review routine alongside the forecast so buyers can respond when actual demand differs from the plan.

ELECTE connects sales, stock and promotion data, compares six forecasting methods on the available data, and presents forecasts, accuracy and exceptions in visual dashboards that teams can share. For retailers assessing demand forecasts alongside their own sales data, visit [ELECTE](https://www.electe.net).
