Startup Valuation with Machine Learning

Inspiration

Valuation of startups has been choppy in the past few years especially in this tumultuous economy. The problem of valuation of startups came up in the panel discussion on the topic of investment landscape, that inspired us to leverage on ML to solve this, as in today's data-driven world, ML has the ability to provide a more objective and efficient way to analyse and predict the value of startups. We hope that our tool will be able to assist investors and venture capitalists in making informed decisions about startup investments.

What it does

Our project aims to create a more reliable and accurate way for startups to place a valuation on themselves, by using a reliable ML model to predict the valuation of the said startup. The accuracy of the valuation will of course be backed by the reliability of ML.

It will have a API endpoint which takes in parameters that will be used for ML model prediction.

How we built it

1) We conducted research with reference to a research paper: Valuation of Startups: A Machine Learning Perspective. It discussed how 3 different types of models are used to predict the valuation of the startups, including multilayer perceptrons, Domain Adaptation Neural Networks and CatBoost.

2) We sourced initial data through kaggle https://www.kaggle.com/datasets/justinas/startup-investments from Crunchbase 2013, cleaned and preprocessed the data using Pandas.

3) We built our machine learning model with the XGBoost model as gradient boosting methods were proven previously in the paper to give optimal results for this particular use case. Data is divided into training and test sets with a split ratio of 80:20. The model is created through Jupyter notebook in AWS SageMaker and deployed into the AWS S3 bucket for us to retrieve to use on the API.

4) By utilizing AWS Lambda, we were able to create a serverless environment for the endpoint generated using AWS Sagemaker. Furthermore, we utilized AWS API Gateway to establish a REST API. When a POST request is received, the API Gateway triggers the corresponding Lambda function, which houses the logic for making predictions using our model. The API Gateway then returns the output from the Lambda function to the client, providing a smooth and efficient process.

5) Next, to make the Machine Learning model scalable, we leverage the auto scaling feature in AWS so that multiple investors can access the endpoint with low latency. We have decided to perform a theoretical calculation and derived an assumption that there are around 125,000 active investors worldwide every minute which equates to 2,000 requests per second.

Challenges we ran into

One challenge we ran into was the cleaning and preparing of the data. Much of the data available to us was actually unusable because the key information, the actual valuation of the startup, was not available. Hence, we had to train our model based on a limited knowledge base.

Another challenge that we faced was troubleshooting in the cloud environment as it is not as straightforward as debugging in the local environment. Fortunately, we managed to make use of the AWS CloudWatch logs to debug for errors and fixed the errors successfully.

Accomplishments that we're proud of

We are definitely proud of having achieved a relatively low error rate of 0.13, which signifies that our model was effective in predicting the valuation of the startups. In addition, we also managed to integrate multiple AWS cloud services into our Machine Learning solution.

What we learned

Our biggest takeaway was definitely understanding the process of creating a machine learning pipeline, from research stage, ideation stage, data cleaning stage, model building stage to model deployment stage.

What's next for Startup Valuation with Machine Learning

From this project, we wish to be able to utilize improved data to train an even more reliable model that can predict startup valuations based on more than the disclosed funding amount from a given funding round. Machine learning has a multitude of applications in the fintech industry and we wish to be able to ideate more great ways to use it as such.