Traditionally forecasting is done as manual work based on expert knowledge. However, this can take a lot of effort and sometimes there just is not enough information for anything other than guessing. Using automatic time series forecasting algorithms allows us to generate forecasts based on patterns hidden in the historical data. In this way, high-quality forecasts can be generated automatically and quickly even for massive datasets.
The required data can be read from your planning system, processed by the automatic time series forecasting, and then written back to the original source or elsewhere. The human expert can then adjust these forecasts if additional information is available.
Forecasting often deals with time series data. A time series is a series of datapoints in time order such as monthly revenue or daily sales. Time series often contain trends and seasonal variation. Occasionally, these patterns are obvious and detectable using naked eye judgement alone, but frequently they are obscured by incidental noise. Automatic time series forecasting can find and analyse such patterns.
Algorithms are often more accurate at predicting outcomes than human experts. Let us assume that you have a thousand time series that you need to forecast monthly two years into the future. With such a large amount of data the likelihood of human error increases and the job is hardly inspiring. There could be a temptation to just copy and paste last year’s data across. Sometimes this can be because there’s not enough time for a detailed analysis but sometimes it can truly be the best possible forecast. Surprisingly often the best predictor is even simply copying the latest observation on and on. Yet often a more complex mathematical model is needed to create the optimal forecast.
The problem is that it is not easy to say what is the best choice and determining it takes a lot of work. However, an algorithm can quickly train multiple models based on historical data and run tests to compare their accuracy. Based on these tests the algorithm decides, which approach to take. Thus, by using such an algorithm you get a detailed analysis for every time series every time you run the forecasts.
It is important to note that there are circumstances in which additional expert input remains invaluable. You might already know, for example, that one of your two production lines will need to be shut down in a year’s time. This information will need to be manually inputted because the algorithm will not be able to pick it up from the historical data available to it. An algorithm is a tool that experts can use to generate better forecasts, faster than before, allowing you to make more productive use of your time.
Automatic time series forecasting can also be used to exploit known interdependencies between different variables. Additional time series based on forecasts drawn up by human experts are a classic example of this. Let us say you want to forecast energy consumption at a production facility. A time series on your planned production volumes can be added to the algorithm. This allows the algorithm to consider the historical data presented to it while also taking account of the variation you are expecting in your production volume.
In some circumstances, even time series that do not directly have historical data can be forecasted. This is an option if historical data is available through other relevant time series. For example, you might have data on past fuel purchases, but you want to forecast your fuel consumption. If you make large one-off fuel purchases but your consumption is steadier, you can use the moving average method to forecast your consumption. This allows you to account for seasonal variation and trends, but the forecast will not be unduly affected by whether the purchase has been recorded at period start or period end.
If the time series you are using contains seasonal variation, it is desirable to have a minimum of two seasons’ worth of historical data. To forecast future monthly temperatures, for example, at least two years of monthly data would be preferable. The more data you have, the better.
Automatic time series forecasting not only helps you to forecast better but also means less grunt work for your experts and more opportunities for actioning the results. Implementing a time series forecast solution is relatively straightforward, and it is an excellent introduction to algorithm-driven analysis.
If you would like some technical information about time series forecasting, check out this article.
We have created a video to demonstrate how our algorithm works in practice. You can view it below.
If you would like to learn more about our advanced analytics and optimization solutions, check out the solution page.