The challenge of moving holidays in demand forecasting
Moving holidays are holidays that occur each year, but where the exact timing shifts from the perspective of the Gregorian calendar system. Chinese New Year (CNY) is an example of a moving holiday, it is based on the lunar calendar. Chinese New Year most often falls in February but can also occur in January. Since the date of Chinese New Year changes from year to year, the effect of this holiday can impact sales in multiple months. It is often the case that production accelerates some time before the start of Chinese New Year, almost completely stops during the holidays, and finally rises to the regular level after the holidays. In these cases, the effect of the holiday is not confined to the seasonal component of the time series since the seasonality rhythm (based on lunar calendar) is not in line with the demand forecast rhythm (based on Gregorian calendar). This often leads to a significant decrease in the performance of the statistical forecast (i.e. lower accuracy and higher bias) for the months affected by the holiday.
Combining statistical forecasting with machine learning
Conventional statistical models (e.g. moving average and exponential smoothing) are widely used within the industry to predict demand. Often with good reason, since these models usually perform reasonably well and they are intuitive and easy to interpret for planners. We have seen, that statistical models (even if we add a seasonal component) are not able to model the complex effects of moving holidays. Based on these observations, we designed a new approach using machine learning to predict ‘uplift’ factors that are used to scale the baseline forecasts. We give a detailed description on how EyeOn addresses this challenge in this blog post: How to make better demand forecasts for Chinese New Year using machine learning
Results case study previous blog
Going beyond Chinese New Year
The results of modelling Chinese New Year for a large multinational are very promising. In this case study, we find enormous improvements in forecast bias. For ten different regions, we achieve an average absolute bias reduction of 68 percent points. We don’t stop there, however! The next step is to extend the logic on two fronts: (1) automatic detection of significant events (rather than manual) and (2) applying the logic to a wider variety of moving holidays in addition to Chinese New Year.
First, we implement automated event detection to determine whether an event significantly impacts a time series in a certain period. If so, we use our machine learning model to apply a correction to the baseline forecast. If not, we just use our statistical baseline forecast. The event detection consists of three statistical tests, which determine if the sales during an event are significantly impacted. The identified events are corrected by predicting uplift factors for the event periods. Automatic event detection results in less manual intervention form the forecaster and provides statistical proof that an event has impact on the time series. Making you less dependent on the forecaster’s ability to interpret and their experience whether a correction should be applied.
Second, besides the uplift factors, the event sales are corrected to improve the forecast performance in the periods after the event. We do this correction to avoid that the sales disruption during the event impacts the statistical baseline forecast. In the case of Chinese New Year, the correction is used to make sure that the statistical forecast level will not be too low after having three negatively impacted sales months.
The extension of this logic makes it possible to better forecast moving holidays and to rely on statistics to determine which events should be corrected. Making it applicable for more customers and events, resulting in improved forecasts during moving holidays. We applied the new extended logic to two other events besides Chinese New Year in a new case study. Comparing the results of the developed method to the results of conventional time series models (e.g. moving average / simple exponential smoothing). In this case, we find an average increase of 13.1 percentage points in forecast accuracy and a reduction of 18.7 percentage points in the bias for the affected months. Note that for the non-impacted months, the two forecasts are identical.
Challenge accepted, challenge completed
Moving holidays can have a significant impact on demand. Moreover, modelling this effect can be challenging since the effect of the event impacts a different period (e.g. week or month) each year. We accepted this challenge and developed a new approach where the conventional time series forecast is complemented with a machine learning algorithm that models the effect of moving holidays. This approach is now available for a large variety of events and customers.
Eager to learn more about this topic or curious if this method could also be valuable for your organization? Please reach out to us!