Some words on Forecasting & Predictive Analytics

In my experience, most companies are good at tracking back data for previous periods and put together an analysis of what actually has happened.

Business Intelligence and Financial Control are entities that focus on telling what has happened, and why it happened. And then the next question arises: what will happen next?

Predicting what will happen next is obviously not as easy as explaining what happened in the past, but, again in my experience, most companies would benefit from putting a little bit more effort into future predictions. However, this should not go that far that it becomes the main point so that forecasts potentially reduce your flexibility and actions. Assume that you forecast your sales to be 45 for the next quarter, but someway during the quarter you suspect that you might be able to hit 60.

What will you do?

Some will have a shot at 70, some re-set their target at 60, but far from all will go for 60. Some deliberately, even.

And if you hit 60, will you set your next forecast at 60, or go back to 45?

But forecasting and future prediction can be a very useful too. It gives the organization the chance to stay alert and plan ahead.

There are two different ways of going about with forecasting: qualitative and quantitative.

The qualitative ones tend to work better in the short-term ahead. They reply on expert or market opinions, for good and for ill. There are two models: market research, where for example a number of customers are asked if they will purchase a product, and the Delphi Method, where experts express their opinions about future.

The quantitative models work better mid- to long-term. They eliminate the human opinion and only look into the data. There are three models: the indicator model, where you look into fairly reliable relationships – for example Product A has had a gross margin of 42% and we assume it will stay at 42% even in future, so what would our sales and cost of goods sold be at given certain volumes; the (more mathematical) econometric model used mostly by macro economists, which examines the data sets over time and tries to find out if there are any relationships; and the time-series model, where data from previous periods are put together to predict the future. There are some variances to this model, which is mostly used by companies, for example you might choose to keep or eliminate certain one-offs, or extraordinary past events, or you might feel that the most recent period should be seen as more important as the most distant period.

Remember that you can not foresee unexpected events or abnormalities in future, for example if you guessed many experts in September 2014 what the crude oil price would be by December same year, many would have probably answered “around $100 a barrel” and not “around $40 a barrel”.  Also, when forecasting, you are working with yesterday´s data, not today, and yesterday was yesterday, so your data is not up-to-date. Also, as we mentioned above, you might look too much into the past and miss out on future possibilities, or your organization might lose out on its flexibility.

Another approach that in some ways is similar to forecasting is the Predictive Analytics, in which you gather a set of data from known periods, find out a pattern that includes parameters such as irregularity or seasonality, in order to help you predict patterns or trends in unknown periods.

Now let’s look at the example below to predict unknown (future) overnight hotel guests stays in a Spanish coastal town by looking into the past data.

 

PA Table 1

 

From the table above, we can read that every Q3 has the highest number of overnight hotel stays, followed by Q2, then Q4 and finally Q1. This is no co-incidence, this is due to the warmer summer months and cooler winter months. Here we can take seasonality into account. At the same time, we also see that there is irregularity in the table. For example, Q1 was the highest in 2010, but Q2 and Q3 were the lowest in the same year. We know that the overnight stays are not evenly phased between the quarters, that there is a specific ranking between the quarters, and that the same quarter might rank higher or lower compared to the same quarter in different years. Seasonality and irregularity are two variables we have to take into account when predicting the future.

This means that you need to use a regression model by first calculating the moving averages, then take into account the above seasonality and regularity, calculating a multiplicative factor for them before de-seasonalizing the data. The final step is to run the predicted overnight guest stays for future periods.

 

PA Table 2

 

When you then create a graph with the “Forecast” column for the quarters, you get the historical quarters together with the predicted quarters. Here the seasonality and irregularity are obvious: Q1 is low, Q4 is high. Also, even though the quarters among themselves have not increased every year, the overalls have, and this is also reflected in the predicted quarters,

 

PA Graph