What Categories Are Worth Doing AI Demand Forecasting For? — Six Data Readiness Criteria and Three Situations to Avoid

AI demand forecasting is not a technical issue of "whether it can be done," but an economic issue of "whether it is worth doing."

When suppliers come to discuss, they often start by talking about how much the model accuracy can be improved. However, what truly determines the success or failure of this endeavor are the seemingly mundane issues before the prediction: how much historical data you have, whether there have been stockouts in the warehouse, whether promotions have been marked, and whether the SKU master data is clean.

This article provides a set of criteria that can be scored on the spot in a meeting: six data readiness criteria, three situations to avoid, and a reasonable expectation for judging "success."

Six Data Readiness Criteria

Each criterion is accompanied by the consequences of failing to meet it. If two or more of the six criteria are not met, don't start with AI; start with data governance.

① Historical Length ≥ 24 Cycles

Criterion: At least two years of monthly data, or more than 24 weeks of data. For categories with clear seasonality, it is recommended to have 36 cycles or more.

Why: The model needs to observe at least two complete cycles to distinguish between "trends" and "seasonality." With only one year of data, the model will treat a one-time seasonal fluctuation as a permanent change.

Consequences of Not Meeting: The prediction will inevitably be wrong at seasonal turning points, and this will be a systematic error, not a random one.

② Data Granularity Matches Decision Granularity

Criterion: The historical data should match the granularity at which you plan to make decisions. If you plan to replenish by SKU × warehouse, you need SKU × warehouse historical data; if you only have national totals, you can only predict national totals.

Why: Accurate total predictions do not necessarily mean accurate predictions when broken down to each warehouse. Aggregation can mask fluctuations, and disaggregation can amplify errors.

Consequences of Not Meeting: The prediction results will be useless for replenishment decisions, and the business department will bypass them.

③ Stockouts and Out-of-Stocks Must Be Marked

Criterion: Historical data should distinguish between "low sales" and "no sales due to stockout." There should be at least one stockout flag field, or the number of stockout days can be inferred from initial inventory + incoming inventory − final inventory.

Why: This is the most common and fatal pitfall in demand forecasting. Sales during stockout periods are not demand, but supply limits. If the model learns from a period of low sales due to stockout, it will assume that the category has weak demand and continue to predict low sales—accurately but disastrously.

Consequences of Not Meeting: The model will systematically underestimate the demand for popular categories, leading to a continuous stockout cycle.

④ Promotions and Events Must Be Marked

Criterion: Historical data should include promotion types, intensity, and time windows; records of major customer orders, new product introductions, and competitor stockouts should also be available.

Why: A sales peak during a major promotion will be treated as a trend by the model. Without marking, the model will predict a non-existent norm.

Consequences of Not Meeting: The model will either exclude or misjudge the peaks, causing the baseline prediction to be off.

⑤ SKU-Level Sales Stability Can Be Assessed

Criterion: Calculate the coefficient of variation (CV = standard deviation / mean) for each SKU. CV < 0.5 is relatively stable, 0.5～1.0 requires caution, and > 1.0 indicates intermittent demand.

Why: Intermittent demand (with many zero sales periods) is not suitable for conventional forecasting models and requires specialized intermittent demand methods or simply not being forecasted.

Consequences of Not Meeting: Including long-tail SKUs in the model will lower overall accuracy, leading to management losing confidence in the project.

⑥ Clean Master Data

Criterion: No duplicate SKUs (multiple codes for the same material), consistent units (mixing individual, box, and pallet units should be convertible), deactivated SKUs are marked, and substitution relationships are maintained.

Why: Forecasting is done by SKU. If one SKU is split into two codes, both time series will be incomplete, and the predictions will be unreliable.

Consequences of Not Meeting: Data cleaning will consume 70% of the project time, leaving no time for business validation.

Three Situations to Avoid

After discussing what should be done, let's talk about three situations to avoid—in these three cases, the ROI of AI forecasting will certainly be negative.

Situation One: High Proportion of New Products

If more than 30% of the sales in the past 12 months come from products that have been on the market for less than a year, historical data is meaningless for these SKUs. In this case, analogous forecasting (using similar older products as references) or small sample methods are needed, not conventional demand forecasting models.

Judgment Method: Calculate the percentage of sales contributed by SKUs that have been on the market for less than 12 months.

Situation Two: High Proportion of Long-Tail SKUs

If 80% of the SKUs contribute only 20% of the sales (a typical Pareto distribution), and the CV of the long-tail segment is generally greater than 1, the benefits of forecasting the long-tail are minimal.

Correct Approach: Only model the A-class SKUs (those contributing 70% of the sales) and use simple rules (such as moving average + safety stock) for the long-tail, which is less costly in terms of human resources.

Judgment Method: Calculate the percentage of sales contributed by the top 20% of SKUs. If it exceeds 75%, focus on this segment.

Situation Three: No Channel Data Feedback

If the data stops at the distributor, and you only have "shipping data" but no "end-user sales data," you are predicting the distributor's purchasing rhythm, not market demand.

Why It's Fatal: The distributor's purchasing rhythm is influenced by rebate policies, financial pressure, and inventory strategies, and may have a low correlation with end-user demand. Using shipping data for "demand forecasting" is equivalent to predicting the distributor's fiscal year.

Judgment Method: Ask, "How much of our data is from end-user sales, and how much is from what we shipped out?"

Three Positive Signals to Look For

Conversely, if three or more of the following criteria are met, the ROI of AI forecasting is usually favorable:

1. Number of SKUs > 500: Manual judgment is no longer feasible, and the marginal value of the model is high.
2. Demand Has Clear Seasonality/Cyclical Patterns: The model can capture patterns that humans might miss.
3. High Cost of Stockouts or High Inventory Costs: A 5% improvement in forecast accuracy can directly translate to reduced inventory or fewer stockouts.
4. Stable Promotion/Major Customer Rhythms: These can be marked and learned by the model.
5. Historical Data Has Been Digitized for More Than Two Years: This is the minimum threshold.

What Defines "Success": Expectations Must Be Realistic

This is the easiest point for suppliers to gloss over. Here is a reference range:

Focus more on bias (systematic deviation) rather than just MAPE. A model with a 20% MAPE and a bias close to 0 is more useful than a model with a 15% MAPE and a +12% bias (consistent overestimation)—because the latter will continuously pressure inventory, eventually leading to stockouts.

The best way to validate is not to look at model metrics, but to conduct a blind test: take the past 6 months, let the model predict and let a human predict, and see whose deviation is smaller. If the model cannot outperform the experienced planner, don't proceed.

In a Nutshell

The success of AI demand forecasting is 80% dependent on data readiness and 20% on algorithms. If two or more of the six criteria are not met, it is better to spend three months on data governance—those three months will yield a higher return than forcing a model into place. The standard for deciding whether to proceed is not "can we do it," but "is our historical data suitable for it."

This article belongs to the "Digitalization and AI Zone · AI Empowering Operations" direction. Related articles such as "How to Calculate the ROI of AI Investment" and "Four Red Lines for Large Models Entering Business Processes" can be found in the zone list.

Knowledge code: 16.3.1 Author: QTank