Final Project

Cindy Pham, Isabel Rodriguez, Lyric Otto

Background

We found a dataset on Kaggle.com containing weight lifting metrics for individual competitors and information about different competitions throughout the world. The dataset contains 900,000+ rows and 37 columns detailing the competitors age, sex, bodyweight, place received during competition, success or failure during lift attempts, IPF/McCulloch/Wilks/Glossbrenner metrics, event participated in and equipment used.

TotalKg lifted is important to know because for most, if not all, lifters, it is essential to know their One-Repetition Maximum or 1RM. A 1RM is a personal weightlifting record for squat, deadlift, bench, or any other weightlifting exercise. 1RM is a precise measure so it helps lifters judge how effective their program is and also allows them to get positive reinforcement and a sense of accomplishment when they can clearly see their progress. With our model, we can predict the total weight a competitor can lift in their competition and see whether the competitor exceeded expectations or has room for improvement.

Exploratory Analysis

IPF (International Powerlifting Federation) points are used to rank athletes in competition more fairly when determining "best overall" lifter in competition. In most competitions, athletes compete in specific gender, age and weight classes, where they are only competing against those in their same classes. By using IPF points, which takes bodyweight metrics, age and gender into consideration, athletes are able to compete together more fairly for the "best overall" award (though these awards are typically still separated by gender). When we look at the top 5 competitors gauged by IPF points, we can see that the top 5 competitors are women, who lift between 799% and 969% of their bodyweight.

These plots show the counts of categorical data throughout the dataset. From top left, we have the total number of male and female athletes in the dataset. Then we have the number of people who chose to use equipment during competition and what kind they used. Next, the age distributions across competitions. Finally, the next 3 rows are dedicated to success and fails of the 3 attempted lifts throughout competition. We can clearly see that the 3rd attempt for all lifts has the highest failure rates, which makes sense as athletes probably get tired during their lifts.

We wanted to see the differences between major powerlifting meets by looking at age, bodyweight, best lifts across SBD (Squat/Bench/Deadlift), and the total kg lifted at each meet. For age distributions, we see lower age ranges in the Collegiate Nationals and higher age ranges in the Masters Championships, which is expected. For weight distributions, we also see lower bodyweight averages at the Collegiate Nationals, it would be interesting to look at gender distributions as well. The Nationals, European and World Championships appear to be very comparable when looking across SBD lifts.

Machine Learning and Our Results

We intended to create a model that predicts the total kilograms that a participant could lift based off of their bodyweight, sex, age, competition placement received and success or failure during lift attempts focusing on lifters who participated in SBD competitions. We omitted the four established scoring metrics because they are directly related to what we were trying to predict and we received very high accuracy with very little effort.

For our baseline, we initially attempted to find the average kilograms that a competitor lifts above their bodyweight. We determined that on average a weightlifter could lift about 5.698 times their own bodyweight by dividing the total kg lifted by a persons’ mean body weight and multiplying that by their total bodyweight. This was a wildly inaccurate baseline. We determined that a better baseline was actually 7 times a weightlifters’ bodyweight. Despite this baseline still yielding inaccurate results, we decided the inaccuracy was reasonable because predicting a competitor's 1RM is a very variable and individual process. This will be explained further in our machine learning process.

After experimenting with several different regression models, we found that a Ridge regression model with polynomial feature expansion of degree 2 performed the best. Though Ridge regression may be an indication of collinearity in our dataset, we believe it is negligible because the majority of the models tested performed at similar error and accuracy rates.

In order to evaluate our model, we decided to look at mean absolute percentage error and coefficient of variation. When performed on the training data, we got a 0.175 for our MAPE (mean absolute percentage error) and a 0.289 for our coefficient of variation. With a 17% error, say if someone lifted a total of 400kg, that means the model predicted a weight that was about 68 lbs off. While at first, this difference is very large, when you consider the context of weightlifting and the columns we had at our disposal, we are decently pleased with our results. In reality, trying to predict and determine how much an individual can lift requires so much more than what we inputted, i.e. bodyweight, age, equipment etc. Each individual is vastly different from one another and there are many other factors that go into how much one can lift such as years of experience, how much training one does, medical history, etc. One probably wouldn’t want to use our model to accurately predict how much one can lift, but it can give a reasonable ballpark estimate to aim for.