Our analysts at Woodseer have been making dividend forecasts since 2011.
Since then a number of patterns have been identified which were codified into an automated algorithm. This is the core of the current Woodseer product and represents our philosophy of solving broad problems with technology and having a small, targeted team oversee the results.
Our automated systems allow us to cover a large number of securities globally at very high accuracy with only a few analysts.
With the advance of machine learning (ML) technology and the increased availability and affordability of large financial data sets, we identified a way to further increase our accuracy without significant additional overhead. This is the genesis of our AI project.
The objective of this project is to improve our forecast accuracy by identifying predictive patterns in financial data.
Our first iteration was intended to identify large dividend increases (e.g. special dividends) and decreases (e.g. through cuts and reductions). Building an AI model means rather than explicitly programming certain patterns into the system we instead show the model lots of data and it identifies the underlying patterns.
We created this ML model using artificial neural networks which, in simple terms, mimic the way the human brain operates. This model has initially been trained with more than 5 years of company financial and dividend history.
The findings from this initial ML model were discussed with our analysts who then researched the securities and made their own determinations. This led to a revision of some forecasts, and feedback for our model.
The development of a robust ML model requires several iterations and substantial trial and error. We have continued to experiment with various datasets and input fields to identify the most relevant information. The current version now forecasts specific dividend amounts for all important securities and these forecasts are compared to our existing Woodseer forecasts.
Our analysts are alerted when the AI model diverges significantly from our forecasts and they are prompted to make a judgement on which is most appropriate.
Since this is a self-learning process, as company dividend declarations are announced, new results are used to refine the model by adjusting the weights of the input data to attain increasingly accurate predictions.
Our data science team continues to trial different datasets and adjust the model to improve its accuracy. Our expectation is as the AI model’s accuracy is proven over time and our confidence in its findings increases then it will ultimately be making our dividend forecasts.
Until that point, the ML piece is a supplemental rather than replacement project. We will be providing ongoing updates (via email, this blog and our company LinkedIn page) as we go forwards.