Prediction of Prices with R (SAGARPA)

SAGARPA (http://www.gob.mx/sagarpa) its a government division that "controls" -between many things- the flow of prices of important products in mexican market.

Depending on many factors, prices change in time generating tendencies of behavior that seems difficult to predict. Such behavior can be seen in the next image, where real behavior is very different than a mean behavior.

In this project I used fiver different methodologies of machine learning to approach and predict behavior of prices. Results are shown in next image:

Above image is divided in two. Left side shows in black line, actual behaviour followed by smoothed function in red and average. Right side predicted behaviour by several approaches.

At first sight seems to be that Holt Winters, a smoothing time series data approach and ARIMA (autoregressive integrated moving average) approach are the best approximations to the real, perhaps because these last are averaging methods, but reality is mostrly unpredictable, and more than the average. In other hand the type and amount of available data is not descriptabling enough, this is to say, we would need to explore more than one variable involved in the dynamic of prices to conclude about prices. But in this case, we could say that Linear model is the worst approch.

As can be seen in results, depending on researcher feelings and financial experts machine learning helps to deduce the future behavior of phenomena as fluctuation of prices.