One of my favorite presenters, John Elder, presented his top 10 analytic mistakes at Teradata Partners.
One of my favorite presenters, John Elder, presented his top 10 analytic mistakes at Teradata Partners.
Lack Data is problem zero – obviously you need data to do data mining and analytics. Without data that is relevant to the problem you cannot use analytics to solve it. In particular it can be hard when there are too few cases to use to train a model (in fraud, particularly, the number of known fraudulent cases can be low). Companies that invest in creating relevant data (by tracking how some high risk customers actually behave when given credit the models did not support for instance) can be very effective and worthwhile.
- Focus on training
Training a model is important but overfit is a big risk. In the end, only the effectiveness of the model against data not in the training sample matters. Sometimes training a model more can make it perform worse, as it is made to fit the training data better and better without necessarily matching other data. Keep some data out of your training set so you can check the model against it later. - Rely on one technique
Any technique can be flawed. Always compare the results of any novel technique to some conventional technique like linear regression as a sanity check. And don’t blame the algorithm for bad results as the modeling technique is rarely the issue – setting up the problem and managing complexity are much more likely to be an issue. So use a handful of good tools as, once the data is ready, more techniques don’t add much to the cost of the solution.
Interestingly, though there are many tools, they share common techniques like decision trees, neural networks, nearest neighbor techniques etc. And while all of them have strengths and weaknesses, none of them outperform an ensemble model based on multiple techniques that just averages several models. - Ask the wrong question
You must aim at the right target, and the right target in business terms. In addition, don’t get lulled by the most accurate model, find the one that matches reality best. Best business outcomes is the only thing that should determine best model. For instance if you were predicting stock prices the model might emphasize smallest error but be happy with always making estimates that were high where a business might be happier with larger errors when those were low (because they profited from a price that was higher than predicted and lost when it was lower). - Listen (only) to the data
The data does speak, and can surprise you, but it is not the only thing to consider. For instance, some data seemed to show that spending less money would improve SAT schools (comparing SAT scores to investment per student in 50 states). But many states have more kids taking the ACT and so those taking the SAT in those states are self-selecting. That skewed the results and finding the problem require thinking about the real-world, not more data analysis. - Accept leaks from the future
Data that is not known at the time of prediction can easily be fed into a model. For instance models predicting interest rates or stock prices can be very accurate if they somehow include data about the trends such as considering the moving average of yesterday, today and tomorrow. - Discount pesky cases
These can mess you up but can be what actually matters. Outliers can be mistakes, caused by bad decimal points for instance, but sometimes the outliers show you what matters (fraud for instance). - Extrapolate
People can fall in love with their models and extrapolate too far. This is particularly a problem for folks working in machine learning who tend to extrapolate from “machines can win at chess” to “machines can think”! - Answer every inquiry
No model can ever answer every question – keep your focus on what the model is for and retest carefully before using it for something else. - Sample casually
If you are not going to use all the data (because there is too much) and when selecting your hold out group (see #1 above) be careful how you select the samples. It is easy to pick biased samples that drive the model in a particular direction. - Believe the best model
Instead build several good models, combine them in every conceivable combination and see which ones work best in combination. And if you don’t have time, just use all the models as more models almost always performs best.A lot of people want to build models that reveal the deepest truth of the universe. But having multiple pretty good models is often more effective.
To succeed:
Mistakes lead to experience which leads to learning and success so be prepared to make mistakes. John tells his students to adopt PATH:
- Persistent – Be persistent and attack a problem in different ways
- Attribute – be optimistic and have a can-do attitude
- Teamwork – bring others to help you
- Humility – so you can learn from others and not expect too much of your technology.