Python in Finance

In Finance, models play a key role in many areas from pricing and risk calculation to planning to automated trading or fraud detection. In this talk, we explore the area of stress-testing, forecasting and decision making and show how python can be used to effectively replace and improve upon traditional non-integrated solutions.

Networks of models

While predictive models are typically developed and tested in isolation, they are eventually used in combination with other models in a complex network of up-stream, down-streamed and cyclical dependencies representing various financial statements, business areas, intra-company functions or entities in the market.

The integrity and consistency of such a network needs to be maintained over time, as individual models get improved or replaced, as business structure evolves and as underlying data change. This is hard or impossible without a programmatic solution.

Therefore models need a clear interface between each other, a simple way of operating standalone and a flexible tooling for integration into more complex networks and use-cases.

Structured approach to model design

To solve the challenges above, we apply the best practices of software engineering and functional programming to define a set of guidelines for model development:

• Pure functions: Models are implemented as referentially transparent functions, fully deterministic and without side effects or access to external systems or databases. That way, models can be re-used in a multitude of scenarios, applications or contexts.
• Interface first: The data flow in- and out of each model is defined first using a global “dictionary” of key/value pairs. A “skeleton” is created, annotated with required and promised data keys and only then incrementally filled with code or broken down into smaller components. This allows individual modeling teams to operate independently and in parallel.
• Continuous integration: Promotion of models code from development to production is governed by continuous testing within increasingly large networks to detect incompatibilities early.

Example implementation

Using python decorators and the dask library, we build a sample model, define its interface and show how it can be programmatically inspected to connect it with other models.

A Jupyter notebook with additional details will be available as part of the talk materials.

Higher-level workflows

By using a decorated function as a basic building block, we compose higher-level constructs such as networks of models and take advantage of “embarrassingly parallel” nature of many applications including sensitivity testing, “Monte-Carlo” simulations or reverse stress-testing, optimization and reinforcement learning.

Role of open source

Last, but not least, we will touch upon the role of Open Source Software in corporations, how the approach has been changing and what individuals can do to help promote it.

Audience

This talk is targeted for

• model developers, not-only from the financial industry (no prior financial modeling knowledge required)
• software engineers working with large systems of models
• users or developers of dask

Keywords

Finance, Modeling, DAG, dask, networks, Jupyter, Excel, SAS