AthlyticZ

▶

▶ 183 ❤ 0 💬 0

50 Days of ML with Sports: Day 7 Your RMSE looks fine. That only tells you the average was fine. It says nothing about whether your model is wrong in a PATTERN, which is much worse than being wrong at random. Three pictures, three assumptions, about four lines of code. Residuals against predictions. You want a shapeless cloud centred on zero. If you see a curve, a smile or a frown, the relationship is not linear and you need a transform or a bendier model. Spread across the range. You want constant width. A funnel shape means the error grows with the prediction, so the model is meaningfully more trustworthy at one end than the other. That does not break your point predictions, it breaks every interval you put around them. A Q-Q plot for normality. Points should hug the diagonal. Two practical notes. Q-Q plots almost always bend at the extremes, because that is exactly where you have the fewest observations. Judge the middle of the line harshly and the tails gently. It is not a failure, it is a small sample. And compute these on held-out rows, not just training rows. The model has already been optimised against its training residuals. Grading those is grading the exam you wrote yourself. #datascience #sports #data

▶

▶ 175 ❤ 0 💬 0

Expectation: hacking into the mainframe while dramatic music plays. Reality: fixing column names, hunting down missing values, testing assumptions, and asking why the model thinks a team will win. The unglamorous part is the real skill. With sports data, a flashy prediction means nothing if the data is messy, the assumptions are weak, or you can’t explain the result. Good data science isn’t just building a model. It’s knowing whether you should trust it: and communicating what it actually means. Which part surprised you most: the cleaning or the explaining? #DataScience #SportsAnalytics #Analytics #RStats

▶

▶ 242 ❤ 0 💬 0

Your model got better. That’s the problem. If you scale your data before splitting into training and test sets, the scaler has already seen information from the test set. That information leaks into the training process, making the model look better than it really is. The safe order is: 1. Split the data. 2. Fit the scaler on training data only. 3. Transform both training and test data using that fitted scaler. Split first. Always. #MachineLearning #DataScience #Python #ModelValidation