Forecasting is the process of making predictions on future trends based on past and present data. Applications range from predicting retail demand and the supply of electricity from solar panels to capacity planning. Any data that includes a time component can potentially be predicted. Let’s face the future, together.
Forecasting is a supervised machine learning method, meaning that it needs a dataset of historic data with known labels for it to learn future predictions. For time series, there are two things that you need to look at: seasonality and trend. Seasonality describes predictable changes that recur every period (for example, every year). Coming back to the solar panel output prediction example, yearly in the Summer months, the output will be higher than in the winter because of the increase in sun hours. Next to that, ever since 2019, the yearly amount of hours of sunshine in The Netherlands have increased slightly. This is the trend that we expect the data to follow.
For forecasting, we typically use Prophet, a forecasting procedure written by Facebook’s Core Data Science team. Prophet takes into account yearly, weekly and daily seasonality, and even considers holiday effects in its predictions. At the core of each prediction, there are four variables: growth, seasonality, holidays and error. The growth function models the trend between changepoints. Changepoints are moments in time when the data shifts direction. The seasonality function models the cyclical patterns in the data, for example, the seasons of the year. Then, the holiday function takes into account a list of public holidays (or other important dates defined by you) and adds or subtracts value from the forecast on these dates depending on historic changes in the values in previous years on those holidays. Finally, the error function takes into account the deviations and changes that needed to happen and were unexplained by the other 3 variables, derived from the training process.
Imagine being able to see what is coming and what will your business. From product demand, to number of passengers or (a lack of) green energy, these insights allow you to respond more adequately to your environment and be more prepared and flexible for changes. If you are tracking time series data, be it event triggers via Google Analytics, or sales with a sales date attached to it, we highly recommend looking into at least some kind of analytics solution. It can’t hurt to confirm your assumptions or be surprised by some hidden insights in the data.
If you don’t collect data yet, getting started with recording time-based events is the first step. Depending on your business, it can be as simple as making sure that receipts or sales confirmations don’t get lost in your email inbox, but get stored in a centralised data storage solution.
Trying to visualise your data is a good next step. If your dataset is not too big, you can get started with opening up Excel and visualising your target variable on the y-axis and your time period on the x-axis. If you already see some kind of trend, or repeatability, there’s a big chance that you’re going to be able to perform some kind of analysis or prediction of the future.
The next and final step is as simple as filling in the form on this page and leaving some information about what you’re trying to predict. We’ll get in touch with some pointers and help you out.
Your message is in good hands. We strive to get back at you within one working day.
