Pitfalls in Systematic Model Development

Joaquin Narro and Monica Caamano

The wide array of considerations made in the book so far should give the reader a feeling for what can go wrong in systematic model development. We will now focus on the two most common mistakes in this area: trying to monetise a model that is just good at explaining (but not capturing) the trend; and underestimating model correlation. We illustrate these two pitfalls with a realistic German power model.

A TREND-DESCRIPTIVE MODEL

We have built a German power linear regression model using 1,400 points of fundamental price data starting from 2014. Our dependent variable is the price of German power (EEX front year future), and our independent variables are the prices of oil (ICE Brent front month), emissions (ICE EUA front December), gas (ICE Title Transfer Facility, TTF, front year), coal (ICE Rotterdam front calendar) and the EURUSD currency pair (foreign exchange, FX).

With these five independent variables, we have obtained an R-square of 0.983 – ie, we have managed to explain 98.3% of the variance. An in-depth look at the regression statistics, shown in Table 12.1, reveals the significance of every independent variable. Out of the five variables, the most significant is