Journal of Computational Biophysics & Chemistry

R-Squared (Lesson 6 of 24): R-squared is one of the most commonly used metrics to measure performance. But it took me 2 years to figure out mistakes that were killing my regression models. In 2 minutes, I'll share how I fixed 2 years of mistakes (and made 50% more accurate models than my peers). Let's go: 1. R-squared (R2): Is a statistical measure used in regression models that provides a measure of how well the observed outcomes are replicated by the model, based on the proportion of total variation of outcomes explained by the model. 2. Range (0 to 1): R2 ranges from 0 to 1. A higher R2 value indicates a better fit between the prediction and the actual data. For example, an R2 value of 0.70 suggests that 70% of the variance in the dependent variable is predictable from the independent variable(s). 3. Mistakes I made that took me 2 years to fix: I made 2 mistakes early on when I was modeling and in particular making forecasts (where the model developed would predict the next n-steps in a sequence). Mistake 1: Using in-sample metrics. Mistake 2: Using Variance explained only. Here's how I fixed them. 4. Mistake 1 (Using In-Sample Metrics): In stats, we're taught to develop a model with a high R-squared or high adjusted R-squared. So what did I do? I tried to increase R-squared of the model, which is produced from the training data. This led to overfitting. My forecast models did not generalize well, and were erratic. The solution: I switched to Cross Validation. This helped, but something was still missing. 5. Mistake 2 (Maximizing Variance Explained isn't Always the Right Solution): R-squared is a measure of variance explained. Meaning that the more the prediction varies with the actual data, the higher the R-squared. This is not necessarily desirable in Forecasting. Rather, in forecasting I want a model that is stable around the mean of the unseen data. To get this, I need a different metric. 6. A better metric (RMSE): Combining R-squared with Root Mean Squared Error (RMSE) was the solution. I would measure both R2 and RMSE for my forecast models, and this gave me a better picture of the model performance. If the model did well on RMSE, I knew the model had high stability. If the model did well on R-squared, I knew the model did well predicting the peaks and valleys (variance explained). And that's how I began making forecast models that were 50% more accurate than my peers. === Ready to learn Data Science for Business? I put together a free on-demand workshop that covers the 10 skills that helped me make the transition to Data Scientist: learn.business-science.io/fr… And if you'd like to speed it up, I have a live workshop where I'll share how to use ChatGPT for Data Science: learn.business-science.io/re… If you like this post, please reshare ♻️ it so others can get value.