In this article we are going to define forecast accuracy, why it is important, how to use the right metrics for each case, what factors can influence it and how it can be monitored.
Forecast accuracy, which is also often referred to as forecast accuracy, is a process within the supply chain whose main objective is to facilitate decision-making and the way in which they impact inventories, purchases, production, transportation, financial margins and customer service.
Its value does not lie in determining a number itself, but in how that number can help make important decisions. If a highly accurate forecast doesn't help reduce stock failures or reduce excess inventory, it's useless.
Forecast accuracy measures how accurate the forecasts generated are in relation to actual sales. This calculation allows us to evaluate how well the planning process is working and allows us to improve future supply and production decisions.
Its purpose is not only representative, but it must be useful for the operation. For example, 90% accuracy may sound like a good thing, but if the error occurs on a high-selling product (a major SKU), it can cause stock failures, customer discomfort and a drop in sales. In short, what is important is not the absolute value of precision, but its real impact on the operation.
On a daily basis, accuracy is measured as a balance between forecast, reality and the costs associated with the error. A forecast with a low level of error that does not capture market trends or changes in promotions may be less useful than one with greater absolute error but greater accuracy.
A well-executed forecast accuracy can be beneficial at different levels of production:
Having an accurate forecast generates internal efficiency, providing competitiveness and sustainability. Fewer planning errors result in fewer resources being wasted, greater responsiveness, and improved market position.
There is no single formula that can be used to define if a percentage of accuracy is “good”. The appropriate precision will depend on the context:
Let's look at a practical example:
Both calculations are correct, but the first is for raw material purchases and the second for making replacement decisions in the store.
The factors that cannot be lost sight of when calculating the forecast can be of two types, depending on how much influence the company may have on them:
Internal factors
These are the factors that are under the company's control and can be most easily monitored.
The accuracy of the demand forecast is also conditioned by elements that are beyond the company's control. Among the most relevant:
All of these factors create uncertainty that cannot be eliminated, but that can be managed with constant monitoring and timely adjustment.
Forecast accuracy is generally expressed as 1 — error (%), but to make the measurement accurately, different types of specific metrics are used. Each metric offers a different way of viewing the forecast and must be chosen considering the objective of the analysis. Let's look at the main formulas, how they are calculated and how their results can be interpreted.
This formula measures the average percentage error in relation to the actual sale and is useful for comparing accuracy between products with different sales scales. However, this calculation can artificially increase the error in products with low turnover, and it doesn't work if there are no sales.
A representative example may be:
This formula measures the error as a percentage of the total sales volume, but it takes into account the importance of each SKU according to the level of sales, so it is weighted, since it weighs the level of sales of each product. This calculation reflects the economic impact since it gives more relevance to the products with the highest sales volume.
For example:
This formula indicates if the forecast tends to overestimate or underestimate the demand.
It is a useful calculation for identifying systematic biases in planning, but it can hide major errors when compensating (over and underestimates).
An example might be:
This formula measures the average error based on absolute units, no matter if it was overstated or underestimated, it is ideal to apply to specific SKUs because it is expressed in the same units that are sold, but it does not differentiate between large and small errors, they all have the same weight.
An example of this case:
This case measures the magnitude of the error, giving greater relevance to large errors, and is useful when errors generate very high costs. It is very useful when it comes to perishable products or products with expensive storage. However, it can be difficult to interpret because it is not in percentage but in units.
Example:
It is very important to be able to define At what level are metrics calculated:
It is crucial to choose in an appropriate way the formula to be applied according to the business process that you want to optimize.
The choice of one metric over another will depend on the type of product, the objective of the forecast and the level of decision where you want to apply it. Below we will see what the main metrics are, how to use them correctly and some examples to understand in which situations each one should be used.
The ideal is Don't rely on a single metric, but rather combine them according to your objectives. Some possible combinations include:
Often the low accuracy is due to How is forecast measured, how is it interpreted and how is it connected to processes. These are the most common errors:
Often aggregated metrics are calculated that show a “very accurate” forecast, but that do not reflect daily reality. As a result, there is confidence in an unrealistic precision, which favors the persistence of errors.
For example: A 5% MAPE at the chain level may sound great, but if the errors per SKU in the store exceed 30%, there will continue to be stock failures.
Sometimes, sophisticated algorithms are sought when the problem is in how the system is configured.
Por example: A forecast with AI can predict 1,020 units, but if the minimum purchase lots are 500 and the presentation stock is 200, the accuracy of the model is obsolete.
When direct comparisons of MAPE or WMAPE are made without considering the horizon, level of aggregation or formula used, misleading benchmarks are obtained that can lead to poor strategic decisions.
For example: To say that “company A has a forecast accuracy of 90% and B of 70%” doesn't mean anything if A measures on a monthly level and B on a daily basis by SKU.
Often the forecast is analyzed as a whole, without distinguishing what part comes from the base model, what corresponds to promotions, what corresponds to events or manual adjustments, etc. This concludes that the symptoms are corrected and not the causes.
For example: a promotion that increases demand by 200% can make it appear that the model failed, when in reality the problem was that it was not reported or it was not considered that there was a promotion.
Sometimes a level of calculation is chosen that does not correspond to the decision that needs to be made, so the metrics cease to be guides for the operation and become reports.
For example: a weekly forecast can show good accuracy for raw material purchases, and not serve to replenish stores day by day.
If accuracy is evaluated without contextualizing external changes, the model is overadjusted for atypical events and performance under normal conditions is affected.
For example: the model cannot be blamed for a 50% error in the sale of masks during a pandemic, when the deviation is due to an unplanned event.
When metrics are calculated once a year or sporadically, errors accumulate and the forecast ceases to be a reliable tool.
For example: A forecast accuracy that seemed stable may deteriorate slowly due to changes in the product mix, but it is not detected in time because there is no monthly monitoring.
Changing the model is not enough if the forecast does not reach the expected level of accuracy. The ideal is identify what is going wrong: the data? , the methods? , integration with the business? , the way to measure?.
Correcting the problems that these faults may be causing is possible. The most effective steps to do this are:
It is important to review the integrity, consistency, and availability of the analyzed data. It can be corrected:
It is necessary to identify if the errors come from internal variables (promotions, prices, changes in assortment) or external variables (climate, inflation, health crises). For example, a forecast that underestimated demand in summer did not have a problem with the model, but temperature was not incorporated as a variable. This can be corrected by taking some forecasts:
Recurrent evaluation is necessary to review whether current models are suitable for each product's pattern. It may happen that a simple model can work for mature and stable products, and fail on items with intermittent demand.
This can be corrected:
It is important to allow the sales and operations team to complement the forecast with insights not reflected in the data.
This is why it is important to have case procedures in which manual interventions are applied to:
It is necessary to synchronize the areas of sales, marketing, finance and operations, etc. so that they all work on the same forecast.
This will prevent that if the marketing team launches a promotion and does not inform the planning team, the forecast fails.
Establishing regular consensus meetings, aligning calendars and having shared platforms in real time help correct these errors.
It is necessary to maintain a continuous measurement cycle and to act by exception rather than reviewing everything every time.
How is it corrected?
When accuracy is low, it's not just about “improving the model”, but about Strengthen the entire system, including data, methods, collaboration and measurement. Thus, the forecast ceases to be an isolated number and becomes a reliable planning tool.
The key is in the continuous monitoring and in translating metrics into real decisions. A functional control system does not search for numbers, but rather to detect relevant deviations in time to be able to correct them.
Some of the most effective practical practices are:
There is no need to measure everything, you have to establish clear indicators according to the business objective.
For example:
Having an automated routine is essential to obtain calculations on a recurring and automatic basis to free up and reduce the number of operational tasks. Ideal values can be defined and alert systems that are activated when metrics fall outside those values.
To implement it you can:
Not all products require the same control, so the effort must be concentrated where the error has the greatest impact.
How to implement it? Build a matrix that combines value and stability, and assign different metrics to each group.
Translating metrics into clear graphics will allow you to identify trends and prioritize actions. A graph will allow you to quickly see:
Adequate monitoring of the forecast will promote learning and will facilitate decision-making. This will cause models to be recalibrated, biases detected early and the team spends less time calculating and more time calculating analyze and correct.
Forecast accuracy is decisive in demand planning, but not an end in itself. Its value lies in being able to use it when making inventory, production, logistics decisions, etc.
What method makes demand forecasts more accurate?
The best results come from combining time series, causal models, AI techniques, and expert judgment of the areas involved.
How to improve the accuracy of forecasts?
With quality data, appropriate methods, inclusion of external variables, collaboration between areas and continuous monitoring.
What are the best KPIs for forecast accuracy?
The most used are WMAPE, MAE, RMSE and BIAS, applied at the appropriate aggregation level.
How to calculate forecast accuracy in Excel?
Applying MAPE, WMAPE, MAE, RMSE and BIAS formulas by period and SKU, aligned with the current forecast at the time of the decision.
How does forecast accuracy impact inventory management?
Higher accuracy reduces stock failures and waste. However, if some parameters are ignored, the improvement may have a limited effect.