Inventory Forecasting Models: MA, ES, ARIMA, Croston & When to Use Each

No single forecasting model fits every product in your catalogue. A steady bestseller, a seasonal jacket, and a spare part that sells twice a year all behave differently, and a model that suits one will mislead you on the others.

This guide explains the inventory forecasting models planners use most: moving average, exponential smoothing, ARIMA, Croston's method, and machine learning. For each one, you'll see how it works, a worked example where the math is simple, and the situations where it breaks down. Some teams call these inventory forecasting methods, and the two terms mean the same thing here.

The goal is to help you pick the best forecasting model for inventory planning in your own business. In most cases that means starting with the simplest option and moving up only when your data and demand pattern justify it.

Why the Choice of Model Matters

Forecast accuracy depends as much on the model as on the data. A simple moving average works well for products with stable sales, but it falls apart when seasonal spikes or sudden shifts in buyer behaviour arrive. Models like exponential smoothing and ARIMA adapt to trend and volatility, though they need clean data and some statistical care.

Pick the wrong one and the cost shows up everywhere: lost sales, excess stock, higher freight bills, and customers who stop trusting your availability. Pick the right one and you get dependable forecasts at SKU or category level, more accurate replenishment, less manual planning, and a faster response when demand shifts. That makes the choice a business decision as much as a technical one, because it shapes profitability and service levels.

Inventory Forecasting Models at a Glance

The Top Inventory Forecasting Models, Explained

1. Moving Average (MA)

A moving average takes the average demand over a fixed number of recent periods. Each time a new period arrives, the oldest one drops out, which is why it's called "moving." The idea is to filter out noise and reveal the central level of demand.

It suits stable products without strong trend or seasonality, such as a fast-moving SKU that sells about the same each month. It's easy to build in a spreadsheet and needs very little data. The trade-off is that it treats every period in the window equally, so it reacts slowly to change and misses seasonal swings. You also have to choose the window yourself, for example 3 months versus 6.

Example: if the last four months of demand were 100, 120, 110, and 130, the 4-period moving average is (100 + 120 + 110 + 130) ÷ 4 = 115 units for next period.

2. Exponential Smoothing (ES)

Exponential smoothing is a weighted average where older data counts for less and less. Recent demand carries the most weight, so the forecast adapts faster than a moving average. It comes in three forms:

• Single ES handles a steady level with no trend or seasonality
• Double ES (Holt's method) adds a second equation to capture trend
• Triple ES (Holt-Winters) adds a seasonal adjustment for recurring patterns

It fits moderately variable demand with a clear level, trend, or seasonal shape, such as promotional cycles, weather-sensitive products, or seasonal demand in fashion. You'll need to tune the smoothing factors (α, β, γ). A higher α reacts faster to recent change, but it also reacts to noise, so poorly tuned settings can make a forecast jumpy. ES also struggles with erratic or intermittent demand.

Example: with α = 0.2, last actual demand of 120, and a previous forecast of 110, the new forecast is (0.2 × 120) + (0.8 × 110) = 112 units.

3. ARIMA and SARIMA

ARIMA stands for AutoRegressive Integrated Moving Average. It's a time series model built from three parts:

• AR (AutoRegressive) uses past values to predict future ones
• I (Integrated) applies differencing to remove trend and make the series stationary
• MA (Moving Average) models the errors from earlier predictions

Add seasonal terms and you get SARIMA, which handles seasonal lags and differencing.

ARIMA works for SKUs with strong autocorrelation or consistent patterns, like monthly demand cycles or high-volume seasonal products. It captures trend, seasonality, and noise in one model, and it's well suited to longer horizons when you have enough history. The cost is complexity, since the data usually needs preprocessing to become stationary. You also have to choose the settings (p, d, q) and compare models with criteria such as AIC or BIC. The result is also hard to explain to anyone without a statistics background.

Example: ARIMA(1,1,1) uses one autoregressive term, one differencing step, and one moving average term. SARIMA(1,1,1)(1,0,1)[12] adds seasonal terms on a 12-month cycle.

4. Croston's Method for Intermittent Demand

Picture a spare part that sells in batches of 15, roughly every third month. Most months show zero sales, then a spike. A moving average or exponential smoothing would smear that spike across the empty months and give you a distorted number.

Croston's method avoids this by forecasting two things separately: the average size of demand when it occurs, and the average interval between demand events. Divide size by interval and you get expected demand per period. That makes it a good fit for spare parts, industrial components, and long-tail SKUs in e-commerce or B2B, where sales are sporadic. It also helps you size safety stock for slow movers and cuts the holding cost of items that rarely sell.

It has limits worth knowing: it doesn't handle trend or seasonality, it assumes size and interval are independent, and it loses accuracy when intervals vary widely. Variants such as the Syntetos-Boylan Approximation correct a small upward bias in the original method.

Example: a part ordered in quantities of 15 every 3 months gives expected demand of 15 ÷ 3 = 5 units per month.

5. Machine Learning Models

Machine learning models learn patterns that go beyond past sales. They can take in price changes, promotions, weather, online behaviour, regional factors, and even competitor pricing. Common families include:

• Linear and ridge regression
• Random forests and XGBoost
• Neural networks, such as LSTM for time series

They pay off in large operations with mature data, especially when outside factors strongly move demand: multi-location retail, price-sensitive seasonal products, or omnichannel inventory. They can capture non-linear relationships and run in near real time as inputs change. They also need a lot of history and context, careful data preparation, and they can be hard to interpret. For a retail view of where this is heading, see AI-driven demand forecasting.

Example: a random forest could predict demand for one SKU using the last three months of sales, marketing spend, weather, product price, and competitor discounts.

How to Choose the Right Inventory Forecasting Model

Start with your products and your data, not the model. Six checks narrow the field quickly.

1. Look at the demand pattern. Stable and linear demand suits a moving average. Seasonal peaks or steady trends call for exponential smoothing or ARIMA variants. Irregular, low-frequency demand, such as spare parts, is where Croston's method is far more accurate.
2. Check volume and velocity. High-volume SKUs benefit from models that respond to fast change, like ES or ARIMA. Low-velocity SKUs do better with Croston or simple reorder rules, since sparse data can't support complex models.
3. Be honest about your data. Moving averages and single ES need only sales history. ARIMA and machine learning need longer series and ideally outside inputs like weather, price, and promotions. If your data is short or noisy, start simple and add complexity as it improves.
4. Match the forecast horizon. For weeks to a few months, MA and ES are responsive enough. For quarters or seasons, ARIMA or machine learning usually perform better, especially when outside drivers matter.
5. Consider your team. MA and ES run in Excel or lightweight tools. ARIMA, SARIMA, and machine learning need a data team or forecasting software, plus ongoing retraining, tuning, and monitoring.
6. Connect it to replenishment. A forecast should drive actions such as purchase orders, safety stock buffers, and alerts. The replenishment quantity formula shows how a forecast becomes an order size.

Most businesses don't settle on one model. They segment the catalogue, often with ABC analysis, and use a different model for each segment. A common mix is moving average or ES for stable, high-volume products, ARIMA layered on for trend-sensitive SKUs, Croston for the long tail, and machine learning where promotions or weather drive demand.

Whatever you choose, test it against real sales and track the error with metrics like MAPE and bias. The guide to forecast accuracy metrics and benchmarks covers how. And since every model above needs history, products with no sales record need a different approach, covered in forecasting demand for new products.

If you're comparing platforms rather than models, see our roundup of AI demand planning tools.

How Supplymint Helps with Inventory Forecasting

Choosing a model is only part of the job. The harder part is running forecasts across thousands of SKUs and many stores, then turning them into buying and replenishment decisions. That's what Supplymint's demand and OTB planning module is built for. Its proprietary AI weighs seasonal trends, weather patterns, yearly events, historical sales, planned markdowns, and closing inventory together, rather than relying on sales history alone.

What that looks like in practice:

• Open-to-buy (OTB) forecasts and buying plans at category and sub-category level
• ML-based automated replenishment that sends stock to stores using actual sales data
• A side-by-side view of actual, predicted, and budgeted sales, so you can see where forecasts miss
• Two-way integration with ERP and WMS, on a cloud platform

Want to see how this works on your own SKUs and stores? Book a demo with Supplymint.

Frequently Asked Questions

1. What Is the Best Inventory Forecasting Model for Seasonal Demand?

Triple exponential smoothing (Holt-Winters) and SARIMA are the strongest choices. Both account for repeating seasonal swings while adjusting for trend and base demand. Simple moving averages and single smoothing usually miss these patterns. For the planning side of the season, see our guide to seasonal demand forecasting.

2. How Do I Choose Between Moving Average and Exponential Smoothing?

Use a moving average when demand is fairly flat and you want something simple. It treats all periods in the window equally, so it reacts slowly. Use exponential smoothing when demand moves around more, because it gives recent sales more weight and adapts faster. Most teams start with MA and switch to ES when the forecast starts lagging.

3. Can ARIMA Be Used for Inventory Forecasting in Retail?

Yes. ARIMA is widely used for SKU-level and category-level demand, especially where demand shows trends, cycles, or autocorrelation. It needs clean, stationary data and some statistical skill, so it suits teams with an established analytics workflow or forecasting software.

4. What Is the Best Model for Slow-Moving or Intermittent SKUs?

Croston's method. It forecasts the size of demand and the interval between orders separately, which beats a moving average or exponential smoothing on items with many zero-sales periods, such as spare parts or long-tail products.

5. How Accurate Are Inventory Forecasting Models?

It depends on data quality, demand volatility, forecast horizon, and the model you choose. Accuracy is usually measured with MAPE, forecast bias, and MAD. What counts as good varies by category, so compare against your own baseline first. Our forecast accuracy benchmarks show typical ranges by industry.

6. How Often Should Inventory Forecasts Be Updated?

Match the cadence to how fast the SKU moves. High-velocity items benefit from weekly or even daily updates. Seasonal or moderate-demand SKUs can run biweekly or monthly, and slow movers monthly or quarterly. Automated systems can update forecasts on a rolling basis, so reorder points and safety stock stay aligned with changing demand.