Stockouts and Overstocks at the Same Time: The Baseline Every Retailer Must Build Before Adopting AI Forecasting

Signals considered

manual forecasting producing high stockout and overstock rates simultaneously
SKU count growing beyond what human experience can cover
forecast deviations significantly amplified around promotions
replenishment plans requiring frequent emergency adjustments

Illustrative scenario. This article explains business-signal judgement and human verification. It does not represent a real customer, conversation, contract, revenue result or conversion claim.

Both Sides of the Inventory Problem at Once

You are the merchandise planning lead. Your company’s inventory management is stuck in an uncomfortable loop: best-selling items keep going out of stock — not because the procurement budget is insufficient, but because forecasts are biased low and safety stock is inadequate. At the same time, some long-tail SKUs have persistently high days of inventory — not because the products are bad, but because forecasts are biased high and purchasing overreacted. Stockouts and overstocks coexist, and inventory turnover keeps deteriorating.

The current forecasting process: each category lead fills out next-cycle demand forecasts by hand, based on their category experience, recent weeks of sales trends, and the upcoming promotion calendar. This approach was barely sustainable when SKU count was low and categories were concentrated. But as the assortment expanded — from hundreds to thousands of SKUs — manual experience could no longer cover planning granularity for every item. Category-level forecast deviations began to diverge: high-velocity items systematically under-forecast, long-tail items systematically over-forecast, and seasonal and promotion-driven items missed by large margins almost every cycle.

The IT team proposes introducing AI-driven demand forecasting. Finance supports it — they want to lower inventory carrying cost. Your challenge: before introducing AI, can you prove how inaccurate the current manual forecasting actually is? Without a manual baseline, you cannot measure AI’s real improvement — and you cannot explain when the AI forecast should be trusted and when it should be manually overridden.

Why Building the Baseline Matters More Than Choosing the Model First

The most common mistake in AI forecasting selection is skipping the baseline and jumping straight to model accuracy comparison.

Without a baseline, there is no improvement metric. Suppose a candidate AI solution achieves a certain accuracy on the test set — is that number high or low? You cannot tell, because you do not know what manual forecasting achieves on the same test set. If manual accuracy is actually close to the AI solution, the huge migration cost — data cleaning, system integration, team training — may not be worth it. Conversely, if manual forecasting exhibits far larger deviations than expected on specific categories or seasons, the AI improvement potential may be big. Either way, you need the manual baseline first.

Without a baseline, there is no trust rule for when to override the model. Even with AI forecasting in production, you keep a manual override mechanism — that is consensus. But when should you override and when should you trust the model? If you know the direction and magnitude of manual forecast bias on promotion items, you can set a rule: when the deviation between AI and manual forecast exceeds a threshold, trigger review; otherwise, default to AI. If you do not know which direction manual forecasting is wrong in which scenario, you have no rule to set.

The baseline has another value: exposing data quality issues at the source. Building the manual forecast baseline requires pulling historical manual forecasts and actual sales for comparison. The comparison itself exposes data problems: were actual sales during promotions compressed by inventory constraints — stockouts making sales look low when true demand was higher? Does the historical data record promotion event timing and intensity? How were new product forecasts arrived at — by analogy to existing products, or by independent judgment? Until you can answer these questions, any AI model is training on compromised data.

Evidence to Verify Before Evaluating Any AI Solution

Before assessing any AI solution, complete these seven data verification steps.

Historical manual forecast accuracy measurement. Pull four quarters of manual forecast values and corresponding actual sales for every SKU. Compute key metrics — by SKU, by category, by time dimension (promotion weeks vs non-promotion weeks, start-of-period vs end-of-period). Do not look at a single overall MAPE — that average can hide massive cross-category differences. Plot the forecast deviation distribution per category — which categories systematically overestimate, which systematically underestimate, which show huge variance.
Stockout contamination of historical sales data. Historical sales data is not historical demand data — if a SKU was out of stock during a period, its actual sales figure is artificially depressed. When building the baseline, mark stockout periods and apply demand-attribution treatment — estimate what sales would have been if the SKU had been available. Without this step, using stockout-depressed sales as demand will cause the AI model to learn “this SKU has low demand” — creating a self-fulfilling forecasting loop.
Promotion event record completeness. Promotions are the largest disturbance to demand forecasting. Check how promotion events are recorded in historical data — start and end dates, discount depth, promotion channel, whether part of a platform-wide sales event. If promotion records are incomplete or formatted inconsistently, the promotion effect on demand cannot be learned by any model. Fix the promotion event data dictionary before training any model.
Seasonality and external factor identification. Different categories have different seasonal patterns — a down jacket’s peak season is the opposite of sandals. Beyond natural seasonality, there are external events — weather anomalies, competitor promotions, category trend shifts. Map the known seasonality and known external influence factors per category. If these factors are not tagged in the data, you will need to add them manually during feature engineering.
SKU lifecycle stage. A new product launched three months ago and a mature product entering its third year have fundamentally different demand patterns. New products are in ramp-up with little history and high volatility. Mature products are relatively stable with ample history. When building the baseline, measure accuracy grouped by SKU lifecycle stage — do not average new product and mature product forecast accuracy together.
Existing replenishment system data interface. AI forecast output must enter the replenishment system to drive inventory decisions. Check the existing replenishment system’s accepted forecast data format, frequency, and interface constraints. If the replenishment system only accepts weekly-level forecasts but the AI solution outputs daily, granularity alignment becomes an implementation obstacle. If the replenishment system only accepts point forecasts and not probabilistic forecasts, the AI solution’s probabilistic output capability is temporarily unusable.
Manual forecaster judgment logic. Before replacing manual forecasting, understand how human forecasters actually make their judgments — what data they look at, what experience rules they use, under what conditions they adjust a formula-generated number. This tacit knowledge is important input for AI modeling. If you replace humans without capturing these judgment patterns, you may lose business causality that no model can learn from historical data alone.

The Human Next Step

With the baseline built, proceed in three stages.

First, use the baseline to identify AI’s entry categories and scenarios. Not every category needs AI — for those where manual forecasting is already reasonably accurate with low volatility, AI’s incremental value is limited. Plot a four-quadrant using the baseline data: x-axis is forecasting difficulty — manual deviation magnitude; y-axis is business importance — sales or gross-margin contribution. Prioritize the “high difficulty × high importance” quadrant for AI — this is where the return on AI investment is highest. For the “low difficulty × low importance” quadrant, keep manual forecasting — these SKUs do not justify the system migration cost.

Second, run a fair comparison on the same data. Use the same time window and the same evaluation metrics, compute manual and candidate AI forecast accuracy separately. But do not stop at overall accuracy — look at category-level and time-dimension breakdowns: on which categories does AI significantly outperform manual, and on which does it actually do worse? Is AI more robust than manual during promotion periods? If AI still shows large deviations on promotion items, the problem may be in promotion data recording quality, not the model.

Third, define manual override rules. Based on the comparison results, design a set of override rules: under what conditions to adopt the AI forecast, under what conditions to trigger manual review, under what conditions to use the manual forecast. The rule parameters — deviation threshold, category scope, time conditions — should be derived from the comparison data, not set by gut feel. After going live, continuously track the frequency of manual overrides and the accuracy change after overrides. If override frequency keeps rising, the AI solution’s applicable scope may have been overestimated.

What Community Messages Cannot Prove

Informal recommendations may say “company X used solution Y with good results,” “an open-source model ranked high in a Kaggle retail forecasting competition,” or “just buy a ready-made forecasting API” — these are directional references, not your selection criteria. Informal recommendations cannot confirm:

How much stockout contamination exists in your historical sales data
Whether your promotion data is complete enough for model training
The actual deviation level of manual forecasting in your categories
How much better the candidate AI solution performs on your data than manual forecasting
In which categories and scenarios the AI forecast is untrustworthy and needs human override

Every item above must come from your own historical data analysis and manual forecasting baseline measurement. The core question in AI forecasting selection is not “which model is the most advanced” — it is “on your data, in your categories, under your operational constraints, how much better than the status quo can AI perform, and better where?” Answering that does not require an AI vendor demo — it requires you to do the manual forecasting accounting first.

This is an illustrative business scenario demonstrating typical evidence verification and decision sequencing in retail AI demand forecasting selection. It references no specific enterprise, AI vendor, contract value, forecast accuracy data, or outcome claim. Actual decisions should follow enterprise sales data, system capabilities, and procurement approval processes.

Frequently asked questions

Is AI forecasting always more accurate than manual forecasting?

Not necessarily. The real advantage of AI forecasting is not that it is always more accurate — it is consistency and scalability. An experienced buyer may be very accurate on the categories they know well, but their judgment does not transfer easily to new categories or markets — and their decision quality fluctuates when they are tired, time-pressed, or working with incomplete information. AI models do not get tired, but their accuracy ceiling is set by the quality and completeness of the training data. If historical data is missing promotion event records or if stockout periods are recorded as low sales rather than as demand constrained by inventory, any AI model will learn the wrong pattern. So the first step is not comparing which is more accurate — it is comparing whose forecast deviations are traceable to a root cause.

Should I always pick the most accurate model?

Not necessarily. Above a reasonable accuracy threshold, explainability is more valuable than incremental precision gains. If a model tells you how many units to stock next month but not why, and the forecast turns out wrong, you cannot tell whether it was a data problem, a missing feature, or a model limitation — you can only blindly adjust inventory. An explainable model can tell you that the forecast was adjusted down by a certain percentage because last year's promotion on this SKU is not running this year — so even if the forecast is off, you can decide whether to manually override. In actual operations, an explainable model with decent accuracy is safer than a black-box model with high accuracy.