Smart FX | Devpost

Inspiration

The portfolio manager no longer wants to waste more time on repetitive and low value-added tasks, namely: rolling futures and hedging Forex exposures, while critical to avoid performance loss due to FX change, are very time-consuming. Convergence of product roadmaps (FX hedge streamlining, AI applications) and other cutting-edge technologies can finally get rid of this very time-consuming part for the end user journey, allowing to laser focus on the current challenges of the Asset management industry: alternative investments, ESG/neutral-carbon alignment, alpha generation, …

What it does

Today, the Forex hedging/rolling process still need human inputs to adjust maturity of contracts, selection of currency, level of hedge ratio, … With our brand-new automatic FX hedging tooling, we propose to leverage our recent developments in AI and UX to infer those more complex parameters, because very context dependent. The solution is composed of the following steps:

Collect time series of FX hedging ratio/exposure, and other impacting factors, for each foreign currency to hedge.
Based on previously built history, implement the most relevant machine learning method to daily determine, if a hedge is needed for each currency. If yes, the target FX hedging ratio and maturity date are also defined by AI.
Capitalize on Fusion Invest APIs to automate the preparation of FX forward orders based on previously inferred parameters and trigger the pre-trade compliance checks
Once compliance is passed (Ok or Breach), the final click to send the orders to the market is kept up to the portfolio manager
A 1-hour daily process is now saved for the portfolio manager (and a few hours monthly for skipping rolling futures)

How we built it

The idea came up after discussion with our existing client base, and brainstorming with different internal profiles: Developers, Data Scientist, Product Management, Sales. Then we built a POC on an internal schema however based on real business case, validated by client-facing colleague, and formerly working closely with portfolio managers. Regarding the Machine Learning part, here are the main steps:

Input (REST/Excel) -> cleaning (Excel/Python) -> model (LSTM) -> Output (Excel)
Implementation of Long Short-Term Memory (LSTM) in Python from TensorFlow.Keras
Model chosen instead of gradient boosting and Facebook prophet based on previous experience, usability, and predicative power.

Each team member brought its own expertise to end up with a packaged solution: developers’ expertise in Fusion Invest hedging tools, data scientists to build data model, architect to integrate systems, product managers to drive the business cases and marketing message, pre-sales to work on great demo.

Challenges we ran into

Identifying the best-of-breed AI/ML method to fit the business need
Collecting data for model training: challenge in terms of volume (long history needed) and quality (relevant business data, inspired from real life production)
Full remote project, with people working on top of their own job activities

Accomplishments that we're proud of

Integrate building blocks coming from internal roadmaps to full new business case
Everyone in the team managed to deliver its own part leading to a great solution
Deliver the first application using AI for Fusion Invest

What we learned

How to articulate several talents into a clear deliverable
How to collect and qualify a critical mass of data to feed the ML model
How to connect initially separated module of the core product, and external solutions (AI/ML libraries)

What's next for Smart FX

We plan to capitalize on this successful solution to all other investment products of Finastra portfolio.
Extend our open APIs to expose the outputs of ML algorithm to build external FX hedging/rolling apps (partner with fintechs)
Generate automatic reports for PM and risk manager
Enrich the model with constraints dealing with trading fees and operational costs
Extend hedging to other market risks: Interest Rate, Inflation, Credit, Volatility