Oh for some Square-esque simplicity in making small international payments. Doing it through a bank is costly and time-consuming. But check out Paypal’s incredibly complex fees policy for international payments, full of gems like this:

For a payment sent from a Norway registered User to a User registered in Hong Kong or China (only where such Currency Conversion Processing Fee is payable by the Norway registered User), 5% above the exchange rate.

As Square well knows, simplicity is valuable and people are willing to pay for it. For consumer international payments, it seems like a lot of money is being left on the table …

Khoi Vinh thinks this new Canon mouse is ridiculous:

As someone who spends a lot of time using Excel, I beg to differ. The mouse is not only a calculator but also works as a numeric keypad for your computer. When spreadsheeting, my right hand is constantly going back and forth between the mouse and the number keypad to enter data. I’m also constantly doing simple arithmetic to check my Excel results. This mouse, if it’s not too uncomfortable, could actually make spreadsheeting more productive and efficient. Although I imagine it might be tricky to get the mouse to stay still when you punch the buttons.

I’ve been playing with Kaggle, a new website that facilitates data analysis competitions. You compete with others to model some data in order to make predictions. I had some fun with the Don’t Get Kicked! competition, which has $10,000 in prizes for the best model that can predict when a used car bought at auction is a dud. After a couple of hours playing with logistic regression models, I’m in 64th place, with a goodness-of-fit score about 11% behind the overall leader.

I think Kaggle is a fun idea, but have some concerns about using it for serious analysis. The competitions are evaluated using a single goodness-of-fit score, like RMSE. This encourages large complex models that fit the data a little better but may be difficult to actually use for forecasting. With regression modelling, adding another variable always improves your goodness of fit a little bit. In the Don’t Get Kicked! competition, the dataset has around 70,000 observations, so there’s plenty of scope to include large numbers of statistically insignificant variables to improve your goodness of fit. I suspect that competition will be won by whoever has a computer powerful enough to estimate a model with all the possible variables and their interactions included.

Of course this is probably the only way that the modelling competitions can be judged automatically. But in real world modelling, other considerations such as statistical significance and diagnostic tests are equally or more important than pure goodness of fit. The Kaggle competitions encourage brute force strength but won’t necessarily result in models that are very useful for those who are sponsoring the prizes. Better to hire an experienced consultant to build you a beautiful bespoke model :)

This is the fourth or fifth time I’ve set up a personal website and blog, and I usually delete them after a year or so. Let’s see how long this one lasts …