Modulai built a deep learning system that predicts a company's annual report one year into the future. By using a single transformer-based model that learns the interdependencies between financial features, it produces more consistent and accurate predictions than separate models for each KPI.
Outcome
1 year
Forecasting into the future
Challenge
The client is a credit score provider focused on the SME segment. Their existing ML models predict individual financial KPIs such as debt, profit, and revenue from historical values. But these fields are related and interdependent, and separate models do not encode those relationships, which sometimes produces erroneous results, such as predicting profit greater than revenue. A model with strong predictive power that learns these interdependencies at the same time would be far more useful.
Solution
Modulai built a single model that learns the relationships between financial features and predicts them autoregressively. A transformer-based architecture, commonly used in NLP for its ability to learn from sequential data, was chosen for the task. The model encodes sequential information about the financial features along with temporal context about the reporting year, which lets it produce coherent multi-field predictions.
Tools
Annual reports for various companies were gathered and used as training data. Different relationships between financial features were specified and constraints on their values were enforced. The processed data was then trained using a JAX-based transformer model. The model was trained on two years of data — the previous year and the current year — and used to predict the financial features of the following year.
Value created
By predicting every field of an annual report from one model instead of one model per KPI, the system keeps its forecasts internally consistent, avoiding contradictions like a profit that exceeds revenue. For a credit score provider, that means a more reliable one-year-ahead picture of an SME's finances to base credit decisions on.
Learn more
