If you use predictive analytics to generate forecasts, you’ve already taken the most important step toward improving planning – congratulations! But what actually makes such a forecast a good forecast?
The initial benefits usually show up quickly: automation saves a lot of time compared to manual planning. Many more forecast objects (e.g., more SKUs, product groups, customer segments) can be processed in less time – not just the top 10% by volume or importance, but the full range from small to large volumes, from blockbusters to new products, from highly volatile to very stable.
But good forecasts alone aren’t enough: demand planners, controllers, and production planners also need to be able to recognize when a forecast is good. This builds acceptance and trust so that forecasts become an integral part of productive planning processes. Here are the 10 most important aspects that characterize good forecasts.
1. As good as or better than a benchmark
If forecasts already exist – for example from manual planning or an existing planning system – it makes sense to benchmark the new forecasting system or method against those existing forecasts. If, for most forecast objects or for a large share of total volume, the new forecasts perform better than the old ones, or equally well but with significantly less effort or in much less time, point to the new system.
2. Achieves small error for the right metric
This is where it becomes clear that what counts as a “good forecast” for one company may not be the same for another. For some, over-forecasting may be less problematic than under-forecasting – e.g., when service level beats inventory minimization and perishability is not an issue; for others, the opposite is true. Forecasts four or five months ahead may matter much more than short-term forecasts – or vice versa. Because relevance differs by use case and process, it’s crucial to define which accuracy or error metric should be used to evaluate forecasts in your specific case. Only then can the system be tuned to deliver forecasts that are optimal with respect to that metric – and that best fit your company. Typically: the smaller, the better.
3. Is reproducible
Consider this thought experiment: on Monday morning, a planner has exactly the same data as on Friday evening – nothing added, nothing removed. They now find the manually created Friday analysis for an item implausible and, unable to reconstruct their thought process, redo the forecast – with a different result, from the same starting point. In research, reproducibility is a quality criterion. The same holds for planning: reproducibility builds trust.
4. Is transparent
Suppose an item exhibits seasonal demand – monthly sales vary with the seasons and holidays, as for gingerbread, swimwear, or agricultural products. The planner knows this. But did the forecast also account for it, or does December still need manual adjustment? A good forecast makes transparent which information went into its creation. Only then can business-relevant facts that were not (or could not be) included – e.g., a recently decided product phase-out or an upcoming regulatory change – be incorporated afterward if needed. Signals you should know whether they were considered include: your own data history, seasonality and cycles, trend behavior, level shifts, associated internal time series (e.g., substitutes, competitor products, marketing spend), and external drivers such as economic or industry indicators, commodity prices, or weather.
5. Knows its limits (part 1)
Intuitively, forecasts for the distant future are usually more uncertain than for the near term. Some methods in combination with certain data are only well-suited for the near future – even if they can technically produce values further out. Sometimes, less is more: a forecast curve that (unfortunately) ends after a few months carries the valuable information that beyond this horizon the forecast might have little substance. This can happen, for example, when working with leading indicators that typically offer a lead of only a few months. Longer-range forecasts may still be needed (e.g., for budgeting), but one should be aware that further horizons generally come with more uncertainty.
6. Knows its limits (part 2): acknowledges uncertainty
By default, a forecast is a single number – a point forecast. The expected high tomorrow in Würzburg might be reported as 13°C. No one will mind much if it’s 12°C or 14°C. Strictly speaking, though, the forecast would be “wrong” – as most point forecasts are. If, however, you acknowledge and report uncertainty upfront (e.g., 13°C ± 1.5°C), then a realized value of 12°C would be perfectly fine. Such lower and upper bounds – typically prediction intervals – better reflect the randomness inherent in real-world processes than pure point forecasts. Wide intervals convey high uncertainty at a glance; narrow intervals (at the same confidence level) indicate higher precision. If possible, include a prediction interval. A good forecast does not suggest a precision it cannot hold.
7. Is objective
Objectivity is achievable, for example, by letting a machine, a fixed method, or a defined process generate the forecasts. This reduces dependence on who happens to create the forecast, and avoids quality fluctuations when different people (due to vacation, job changes, retirement, …) do the work. First, let the data speak – sober and factual. Human strengths enter afterward, when domain knowledge is added that the machine cannot know: special effects such as natural disasters affecting demand, or a planned acquisition of another company.
8. Adapts to new realities
When a new data point arrives, the forecasting method should consider it and, if appropriate, adjust previously produced forecasts. For time series, these are the most recent periods, e.g., last month’s revenue. Updating ensures that new information leads to refreshed forecasts. Most statistical and machine learning methods can learn in this way and produce updated predictions. Some even calibrate how strongly to incorporate new information based on how well that strategy would have performed in the past. Especially in crises, situations can change quickly – adaptability is critical.
9. Is reliable over time
Forecasts for a given time series are usually updated repeatedly. With high automation, this can be done more frequently and more cost-effectively than manual planning of each object. That makes it even more important that you can rely on forecast quality over time. Continuous performance monitoring maintains trust and, through transparency, reveals where strategy adjustments might be beneficial.
10. Knows when it’s good
Finally, the most important aspect. A forecast can be great – but planners need to know that. Only then will they trust predictive analytics outputs and use them directly in planning. Planners should be shown, transparently, which forecasts (across potentially many objects) are truly good, and for which ones expert validation – or even manual adjustments – is advisable. Reasons include generally hard-to-predict series, project business, newly introduced products, or company acquisitions. The goal is for predictive analytics to provide real support so planners can sleep better. A good forecast knows when it’s good – and flags when it isn’t.
Those are our 10 criteria for a good forecast. We hope this provides a useful checklist so you can answer what “good” means for your specific application. Forecasting doesn’t have to be complex. Often, simple tweaks deliver huge value over the status quo. Today’s demand planning tools and software can significantly support planners and make their work easier. The key is to get started – and now you know what to look for.