Here is a graphic representation that will help you understand what predictive analytics is and the value it can bring to your company.
As usual, we will initiate our article with a simple example that allows us to define predictive analytics.
Let’s use a chain of retail stores that decided to launch its annual publicity campaign to sell its products. In order to “prevent” the campaign’s impact on its revenues, the company will “confront” the past campaign results with its present ones.
The objective is to study the link there has been in the past between the sales revenue and the publicity budget, by making other indicators on the store intervene (number of families on the sector, area, etc). We then wish to generalize the links observed in new unknown data.
Usually very greedy on resources, predictive analytics requires high competences.
The calculation power of new Business Intelligence software male the predictive models much more accurate, simple and rapid to use by business users. It is not required any longer to have extensive statistical knowledge to produce a predictive analysis.
Dashboards software are now very current applications in terms of predictive analytics in a company. It allows the business user to project him/herself at a short or mid-term by furnishing precise information on the possible impact of specific situations or decision taking. It is a true revolution in the decision-making process.
What can predictive analytics bring to your company and what are its key steps?
Predictive analytics guides you “rationally” in the possibility of choices to orient your strategy from an objective standpoint.
According to the supermarkets’ graphic representation example, the predictive allows making decisions in a rational manner by basing yourself on past data and mathematical models.
From the company’s point of view, predictive analytics offers your non-IT collaborators and non-statistical the power to analyze the trends and the relationships in your data to predict the evolution of business indicators.
The first step (learning) consists of training the model from known statistical methods (linear regression, logarithmic regression, polynomial regression, simple exponential smoothing, Holt smoothing, Fourier’s transformation,etc.) to build the predictive model.
The second step (prediction) allows predicting the results by using the “intelligent” model.
To finish, the last step (decision-making), more business-centered, consists of consistently using the previously obtained data to adapt your strategy.
What are the most effective predictive models and when to use them?
Choosing a certain predictive model depends on the situation as well as what we wish to show with our data. Therefore, if we dispose of data with trends or cycles, r if we want to produce a comprehensible modelization for a large public, we would use different approaches.
We grouped in the following table the different types of data models and the situations in which they would apply best.
Linear regression
Linear regression is a very common modelization tool. It searches for a linear relationship between the measure to predict and the time axis.
The ideal situation to use this model is when the measure to predict is proportional to the time axis. However, it can equally be interesting to choose this model for its visualization simplicity (a line), which makes it rapidly understandable to a large public.
Logarithmic regression
Logarithmic regression has the same characteristics as of the linear regression. The difference resides in the fact that it allows discovering a logarithmic relationship between the measure to predict and the time axis.
Polynomial regression
Polynomial regression is a more complex form of linear regression. It allows approximating a measure not through the means of a line, but by a polynomial.
Simple exponential smoothing
Contrarily to the regression techniques that are not proper to the time series, smoothing takes into account the specificity of the time variable. In fact, the importance to place into a value decreases over time. For example, to predict the 2017 sales revenue, it is probable that we will attribute more importance in the 2016 sales revenue than the sales revenue of 2008. The diverse smoothing techniques allow considering the depreciation of information over time.
Simple exponential smoothing allows to smooth data and to predict the subsequent value. It applies to data not containing trends or seasonality.
Double exponential smoothing
We can ask ourselves why we would use double exponential smoothing if the Holt smoothing is more precise. The main reason being the calculation time. In fact, in the case of Holt smoothing your need to estimate two parameters, against one in the double exponential smoothing case.
Holt-Winters smoothing
Holt-Winters smoothings allow taking into account data presenting trends and seasonalities.
There exist two versions of the Holt-Winters smoothing:
- Addictive version
- Multiplicative version
The addictive case corresponds to seasons which the amplitude remains constant over time. Whereas, the multiplicative case corresponds to seasons which the amplitude increases/decreases over time.
Fourier transformation
The algorithm allowing to predict the measure is not the Fourier transformation. It is a prediction algorithm using the Fourier transformation.
Technically, the objective is to decompose the measure to predict into a sum of sines and cosine functions of different periods. Thanks to it, our algorithm is capable to recognize more complex cycles and seasonalities than the ones presented in the Holt-Winters part.
Predictive analytics in the Business Intelligence software DigDash Enterprise?
We estimate today that each business user should be able to produce predictions based on data. So, we decided to integrate, since July 2016, predictive analytics as a standard in DigDash Enterprise.