Donaldson Capital Management (DCM) has been an Investment Advisory firm for almost 20 years. With a total portfolio of over $600 Million and historical performance that beats the S&P 500 index, DCM knows that it takes data-driven decisions to be successful in the investment business. Greg Donaldson, DCM's founder, has been using BigML to analyze S&P 500 companies' performance and he came to some interesting conclusions. "Dividend growth is a better predictor of winning stocks if the growth is consistent and persistent. [...] Over the last 5-6 years, a time as chaotic as any we have seen since the 1930s, the winning recipe in the stock market has been dividend-related." This analysis has confirmed the dividend centric investment strategy already in place at Donaldson Capital Management. DCM has many data analysis tools, like Bloomberg and Value-Line. But using BigML on top of these tools presents new insights and enables new data mining possibilities. You can read more on the insights Greg found.

When car accidents happen, a lot of data is gathered, especially when the crash ended in a fatality. The US National Highway Traffic Safety Administration keeps track of these accidents and keeps a ‘Fatality Analysis Reporting System’. This system is a rich source of data for all kinds of analytics. We have transformed some of that data into a more readable dataset. We used it to create a series of models. ‘Car accident injury types’ shows you the injury severity of the various participants in an accident. We have made the prepared data available for you at the Windows Azure Data Marketplace.

Your model is based on the training data you supplied. As you can expect, it will do a decent job predicting the cases that you fed into it. Overfitting occurs when a model does too good a job: it branches out into very specific cases that fit your training data well, but that are useless when you use the model on new data. Pruning is a technique used to prevent overfitting. Common practice to know if your model is overfitted is to test (or evaluate) it with data that you have not used for training. More on overfitting and other machine learning basics in “Everything you wanted to know about machine learning, but were too afraid to ask” part one and two.