John-Paul Clarke, Chief Science Officer


June 12, 2017

Musing on Prix science

I often get asked “What makes Prix special?” My answer is that, apart from the fact that we have an incredible team of people, the thing that makes Prix special is that we do not assume that history will always repeat itself. What do I mean?

Most analytics and machine learning techniques rely on historical “trends.” Specifically, you look at large amounts of data, search for a correlation between the values of two or more variables, and then, assuming this correlation still holds, use it to predict the value of one variable based on the value of the other variable(s). The simplest such correlation would be a linear least-squares fit, i.e., the straight line fit through an x-y scatter plot of variable x and variable y. Of course, there are much more complex correlations in use, but in the end, you are essentially fitting a model to historical data.

At Prix, we also employ a proprietary technique that allows us to identify the “underlying rate” of any arrival process that has the properties of a Poisson process. What do I mean? Basically, we get a very good estimate of the “true” arrival rate of the Poisson process with only a few observations, and can thereby accurately forecast the number of future events. Without this novel technique, the only way to accurately estimate this underlying rate would be to wait for a long time to observe a great number of events. Why is that useful?

There are many processes in the world that can be accurately modeled as a Poisson process, i.e., as a process where the statistical distribution for the interval between successive events is an exponential function. For example, the bookings for a specific room-type, i.e., normal room with a king-sized bed in a hotel can accurately be modeled as a Poisson process. So too can the scoring in a basketball game or the requests for mobile phone repair services. Imagine how useful it is to be able to tell the total number of bookings for a specific room type on a specific day next year, after the first 30 to 60 days of the sales period. Or tell the final point score after the first 15 minutes of a basketball game.

Further, if you think about the historical trends I spoke about earlier, you will quickly conclude like we did, that our proprietary technique also allows us to tell when the historical correlations that have been developed are no longer valid. Thus, we get the best of both worlds. We build the historical correlations like everyone else does, but we also use real-time data to determine when these trends are no longer valid and how to adjust them.

Related posts

Validating Sibyl Forecasting

A comparison of Sibyl, the underlying engine of Prix, to other time-series forecasting methodologies.

Read Post
John-Paul Clarke, Chief Science Officer
02 February 2017

Results from M3 forecasting competition

Two key questions for any person or group doing forecasting are: (1) How good are we in predicting reality? (2) How good are we compared to others?

Read Post
John-Paul Clarke, Chief Science Officer
07 July 2017