| Abstract |
This thesis provides insights into a new method that would analyze and predict the attractiveness of investments in startups for different investor types. Currently, the field is experiencing an increased number of predictive tool development. This thesis provides another method that combines previous research-validated algorithms and generalizes the results.
There are different challenges that investors (the main beneficiaries of the thesis) face in predicting the attractiveness of investments: the availability and accuracy of the data, personalized prediction algorithms and workforce resource limitations. To address these challenges, this thesis aims to identify and evaluate the most efficient prediction models. It is important to keep in mind that investors in most cases will not be able to get the most accurate and standardized data about startups and their goals, which limits the success of any model.
Three multiclass multilabel algorithms were created - Deep Neural Networks (DNN), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM). Important to note that the implementation for XGBoost and SVM was done with OneVsRest algorithm. A startup dataset up to 2013 from Crunchbase, an investment discovery platform, was used as training data. The training dataset contained almost 200 thousand companies, with 40 features and 80 thousand investments with 10 features. Custom evaluation metrics were created to train and improve the model implementations. The result (output) of these algorithms was the probability that a specific investor would invest in the selected company. This output also encapsulates the fact that combinations of investor types could invest in a company.
All three models demonstrated high accuracy rates exceeding 97%. However, the analysis based on F1 scores (around 35%) revealed a substantial issue with false negatives, indicating difficulties in predicting minority classes accurately. Among the models, XGBoost (training time over 5 seconds) and DNN (training time over 3 minutes) showed promising results in terms of training efficiency and overall performance; however, the SVM (training time over 29 minutes) implementation was 3 times better in predicting the minority classes.
For investors, investment pre-screening is one of the most resource intensive process, still mainly happening manually. Therefore, prediction algorithms for investment attractiveness help investors in making more informed decisions at scale. The benefit being the efficiency of pinpointing high-potential opportunities and optimizing resource allocation for better returns. Future research should focus on improving dataset quality, addressing class imbalances, and incorporating more data sources for richer insights. Efforts should also aim at reducing false negatives and developing practical applications for easy integration of these models, ultimately making predictive analytics a standard tool in investment pre-screening.
This thesis consists of 69 pages, 12 figures, 2 tables, and 2 appendices and has overviewed 83 literature sources. All models were run locally on Apple MacBook Pro M3 2023 series computer using Jupyter notebook, GPU acceleration was not used. |
