FutureMetrics: Using Deep Learning to Create a Multivariate Time Series Forecasting Platform for Economic Strategic Planning
- Industry and Business Use Cases
- Financial Services
- Moscone South | Upper Mezzanine | 159
- 35 min
Liquidity forecasting is one of the most essential activities at any bank. TD bank, the largest of the big Five, has to provide liquidity for half a trillion dollars in products, and to forecast it to remain within a $5BN buffer. The use case was to predict liquidity growth over short to moderate time horizons: 90 days to 18 months. Model must perform reliably in a strict regulatory framework, and as such validating such a model to the required standards is a key area of focus. While univariate models are widely used for this reason, their performance is capped preventing future improvements for these type problems. The most challenging aspect of this problem is that the data is shallow (P>>N): the primary cadence is monthly, and chaotic nature of economic systems results in poor connectivity of behavior across transitions. Beyond the basics (regularization, early stopping, etc.), the focus is on model architecture to tackle this challenge in a robust way. Specifically, we will look at a new network design to incorporate stationarization, simultaneously forecasting both growth and totals for every metric. Other areas of technical focus:
1) New tools for introspecting deep learning models.
2) Transitioning from Use Case approaches to MLOps, requirements and roadblocks
3) The use of Granger Causation for feature selection
4) How OOP templates and agile can massively accelerate your Data Science Use Case productivity even when most of your team doesn’t have object oriented programming knowledge.
Goal is to create an MLOps platform for these types of time series forecasting metrics across the enterprise.