Predicting Kickstarter Success

In this predictive analytics project, I've developed a model to predict the success or failure of a given Kickstarter fundraising campaign using a gradient boosted random forest classifier

Context & Summary of Results

Kickstarter is a crowdfunding platform with a focus on bringing creative projects to life.

As part of a data mining course, I was tasked with developing a classification model which would be evaluated solely on the accuracy of its predictions on an unseen dataset. The model could only use attributes which would be known at the time a project is published on the platform.

Upon evaluation, my final CV score was 72.1% - near the average of 72.5% across 84 analytics students. The maximum accuracy achieved was 76.6%. During my own model creation process, I achieved ~72% accuracy during training and ~73% accuracy on a test set. Overall, this demonstrated the appropriate balance between bias and variance that is characteristic of a stable and generalizable model.

Process

At a high level, these are the steps I took in developing this model. This repo currently has the code for my final model. I hope to upload some of the behind-the-scenes code while developing that model once I've gotten that cleaned up.

Preprocessing

• Removed irrelevant variables, including IDs, time stamps, and variables not known at project publication (e.g., staff pick)
• Computed a new 'goal_usd' variable which expressed the goal in the same currency as other key monetary variables
• Binned categorical variabels with many unique values according to domain knowledge
• Eliminated multicollinearity among variables
• Eliminated outliers in terms of fundraising goal and days from project creation to project publishing on the platform

Feature Selection

• Used Random Forest feature importance scores to guide initial feature selection
• Noting low feature importance for most categorical variables, used binning to reduce the number of unique values and enhanced their importance in the model
• The variable for deadline year met the feature importance score threshold but was later dropped due it observed negative impact on model accuracy

Model Selection & Hyperparameter Tuning

• Tested bagged random forest, gradient-boosted random forest, and artificial neural network models on the selected features
• Selected gradient-boosted random forest as it achieved the highest accuracy score in k-fold cv tests
• From GridsearchCV, the best parameters were identified as a 0.05 learning rate, a max depth of 3, a minimum of 2 samples per leaf node, and 100 trees