Increased value creation

With simple statistical forecast models


December 2017

Would you like to read the article in German?


Thomas Georg Nellemann Holm

Complex statistical models often fit past data well, but predicting the future, simple statistical forecast models are often more accurate

The future is never exactly like the past. While complex statistical models can fit past data well, they are dubious when predicting the future. On the contrary, simple statistical forecast models do not necessarily fit past data well, but are, according to many, forecasting scientists predicting the future in a more accurate and robust way.1

Complex supply chains and global markets make forecasting challenging. Some companies choose to accept the uncertainty as insoluble and thus abide with low forecast accuracy. However, it is possible to embrace the opportunities and risks of the future. To do this, we need to be agile and prepared to react to utilise the opportunities and mitigate the risks, “to dance with chance and make luck work for us”, as Spyros Makridakis says.

The solution to this, however, is not to create complex algorithms that are supposed to reveal the future. The belief that we can accurately forecast the future is an illusion of control, and both statistical models and people have been unable to capture the full extent of future uncertainty and been surprised by large forecasting errors and unforeseen events. The biggest challenge of statistical forecasting is complexity, leading to lack of transparency and a lot of frustration in terms of fitting various parameters and trying to understand the outcome. As scientific literature shows, using complex algorithms does not create better results, so why bother?

Simple statistical forecasting methods are, at least for repetitive decisions, typically superior to expert judgements. They are also easy to understand and are therefore easier to add input to. Thus, proactive and robust sales forecasts can be created by combining simple and easy-to-understand statistical forecast methods with the market knowledge of the people in the organisation.

How to build robust accuracy by merging market insights with a simple statistical forecast

Based on our experience, we only need a few simple elements to get a solid statistical baseline forecast. If we are to achieve an accurate and reliable statistical baseline forecast, we need to ensure that the input – in the form of historical sales – is cleansed from significant outliers and events, which otherwise would lead to a biased and inaccurate forecast.

We cannot ignore seasonality, since it can have a high impact on decisions. However, using traditional methods for controlling seasonality might only lead to more complexity, less trust and thereby not achieving the intended forecast accuracy. We recommend using group seasonality logics, which is simple and easy to understand, since it is basically just an index that we add to the constant baseline forecast. We can use this for all products, even NPI, with few or no periods of sales. We accumulate sales history across a range of products, which, due to the law of large numbers, results in a clearer and smoother seasonal pattern with less noise and variance.

Significant step changes of demand create huge challenges for the statistical forecast and can be the source of a lot of manual effort to manipulate history or adjust future sales. We need to handle this in a simple manner. Since statistical forecast is always reactive, it can never foresee step/level change caused by, for example, new listings or customers. Sales needs to provide this information, since statistical forecast needs some periods of observation to “catch up” – a reaction period. The step change logic resets the forecast at the right level and thus achieve a better and more accurate forecast. Incorporate sales insights via tools and templates to establish a clear and structured sales forecasting process and simultaneously minimise optimism bias and document assumptions. Insights from Sales and Marketing should exclusively be integrated at the level that matches the decisions that the forecast must support and can be integrated with a reduced workload by using segmentation that endorses a focus on high-impact products.

The best medium-term sales forecast with the lowest bias and the highest stability is therefore a combination of very simple and understandable statistical forecast and market insights, including opportunities and risks based on simple rules that enables incorporation of market insights into the sales forecast.

Would you like to learn how to do this in SAP Integrated Business Planning? Read more here