How I Built an LSTM Deposit Forecasting Model at SVB UK Using PyTorch
A firsthand account of replacing static deposit forecasting at SVB UK with an LSTM, and why data pipelines and governance became the harder problems.
SVB UK's treasury team previously relied on moving averages and fixed percentage multipliers to forecast deposit flows. However, SVB's deposit base is made up primarily of institutional clients, such as private equity firms and venture capital funds, with irregular deposit amounts tied to funding cycles and market sentiment. After testing macro variables like interest rates and GDP figures, the correlation did not prove useful, and the effort required to manage external data sources outweighed the potential benefits.
The team ultimately simplified their model to use only the previous month's end-of-month deposit balance. They experimented with ARIMA models and various LSTM architectures, ultimately finding that a standard LSTM provided the best Mean Squared Error results and was easier for the treasury team to understand. PyTorch was chosen for its intuitive debugging and compatibility with their iterative research process.
The model's learning involves adjusting weights to minimize the difference between predictions and actual deposit balances, using gradient descent to iteratively improve the model's accuracy.
Written by urgent.news from HackerNoon's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
