When making their operational, tactical and strategic decisions, companies functioning in the retail sector (and in every other domain for that matter) often face the daunting challenge of trying to predict the future as accurately as possible. Businesses want to know whether their sales and expenses will go up or down in the future, whether their inventories will last through the next several months or not, whether their market share will increase or not, etc. Accurate forecasts about future events lead to better policy planning and decision making. How firms go about forecasting future values of interest, largely depends on the available data.
Bring Data into Decision-making
Forecasting methods usually can be described as qualitative (or judgmental) and quantitative. When there is little or no historical data available at all, qualitative methods can be applied. It is worth mentioning that although these methods can be inconsistent, they are not “shot in the dark” guesswork. Using a well-structured approach can lead you to a good forecast even when no data is available. Judgmental forecasting can be an interesting topic in itself but we are going to draw more attention to the quantitative methods of making predictions.
When sufficient historical data is available and the reasonable assumption that the indicator of interest will continue to show a similar behaviour in the future holds, quantitative methods can be used. These methods rest on solid statistical foundations and can lead to very accurate predictions. They can be applied to time series data – collection of data points obtained at regular intervals in time. Naturally, the goal is to predict how the time series values will evolve into the future. How much data is “sufficient” is not easy to say, since it depends, among other things, on the granularity of the data and the statistical method used. For monthly data at least 3-4 full years of data are necessary for an adequately good model to be trained.
Companies that are ahead in their digital transformation can keep track of many different time series. Forecasting a large amount of time series may have been a nightmare in the past, but not anymore! It is easy to plot and look into many time series at once, without having to write many lines of code.
Using modern software and data science packages, different types of time series models can be trained simultaneously for all of the time series that pose an interest. As an example for such packages, we should mention the newly-released tidyverts suite in R. On the plot below, several different time series models with different parameters have been trained and information about how well they fit the data is given.
Apart from making forecasts about future values, using some of the newest statistical software packages (in R, for example) gives you the opportunity to create other plots which can give different perspective over the time series and allow you to better understand the behaviour of the time series under consideration. For example, a seasonal plot can be easily created, allowing the observation of the underlying seasonal pattern and easy comparison between the series. It is also useful in identifying periods when the seasonal pattern changes.
Statistical Learning in Action & Our Experience
We have applied our approach to time series analysis to different use cases like:
- Forecasting future expenses under very high uncertainty
- Forecasting future sales for a very high number of products and estimating future market share
With this newly gained knowledge our clients were able to allocate their budget more efficiently, optimize cash flow management and increase performance in general.
We do hope that this topic was of interest to you and that we have captured your attention.
With this series our aim is to increase data science coverage and to make data-driven decisions an integral part of more companies around the Globe.