Week 5: WAIC up with model comparison скачать в хорошем качестве

Week 5: WAIC up with model comparison

In this episode we compare two models based on two data sets of MooseCo customers. Each is sampled at the same 7 MooseCo sites but one counts customers who took less than an hour to decide to transact and the other, well, more than an hour. We use a simple grid approximation model to mash together data with hypothesis about the average intensity of customer transactions under the two regimes of low and high touch sales experience. We use log predictive probabilitiles (log odds really: lppd, for short) and a volatility penalty (pWAIC = variance(sum(log(probability of each observation, given the model)) to measure information uncertainty (Wide Area Information Criterion (WAIC = -2(lppd - pWAIC)); entropy and the 2nd Law of Thermodynamics at work) and forecast predictability. We then compare the uncertainty and volatility difference between the two regimes and their Wide Area Information Criterion. Watch me fumble through a scatterplot - yikes, the spreadsheet platform was not very forgiving! A reboot of the platform allowed for a simple yet fairly decisive view of two distinct customer decision regimes.