How Riskfuel is slashing compute costs and transforming financial markets with AI

"W&B is the most robust and scalable way to track multiple machine learning projects, without having to handle a lot of infrastructure about tracking or tagging models."
Philippe Chatigny
Senior ML Research Engineer

In the high-stakes world of financial derivatives trading, speed and accuracy are paramount. Banks and trading desks spend enormous amounts on compute power to price complex financial instruments and measure risk at breakneck pace, all while maintaining accuracy.

Riskfuel is a machine learning pioneer in this space. The Toronto-based company is an award-winning leader in AI innovation for the capital markets, helping their clients in financial markets accelerate their models without sacrificing on accuracy by converting traditional valuation software into deep neural networks (DNNs).

“Banks spend a lot of compute power running Monte Carlo simulations to get an accurate value of their risk measures,” explained Philippe Chatigny, Senior ML Research Engineer at Riskfuel. “We train neural networks to be mathematical replicas of these programs and by doing so, help offset this cost significantly.”

A revolutionary solution: Neural network replicas

The problem facing financial institutions is fundamental: they need to calculate accurate expected values (EV) for functions that are inherently stochastic. Traditional approaches rely on intensive Monte Carlo simulations of stochastic differential equations where the accuracy is determined by the number of simulations computed. With millions of dollars hanging on fractions of a percentage point in accuracy, this computational bottleneck isn’t just a technical hurdle—it’s a strategic liability. Banks often spend enormous sums on compute to run these simulations, yet still struggle to achieve the perfect balance of speed and precision that modern markets demand.

“We’re very motivated to solve the core problem: reducing the amount of compute costs in the financial industry, which is really mind-boggling in terms of the amount of energy and money spent within that,” said Chatigny. “We’re really focused both on giving an edge to our clients— because one obviously wants to have an answer more rapidly— but also from an environmental impact too, where we can say we significantly reduce the total compute required.”

Past the computational resources, financial firms also face an operational constraint with their existing pricers. If a trading desk uses daily risk sensitivity measures, they have all night to compute what they need, making sure all parameters are well tuned. But real-time risk calculations are much more demanding. Firms are limited in the accuracy they can get when pricing their options or exotic derivatives in real time.

Riskfuel’s revolutionary approach to helping financial organizations solve this challenge is by building deep neural networks that can replicate the behavior of complex pricing engines with higher level of accuracy than is typically used in production, but which execute much faster. Their DNN models are capable of providing real-time risk metrics at a fraction of the compute, offering traders a significant advantage in their operations.

“By building a mathematical replica of these functions, we can cover any possible range of scenarios that can occur and get better accuracy with equivalent or less compute power and time offsetting the cost of running these functions,” explained Chatigny.

By specifying a domain of approximation to cover all possible market conditions and training a network within this domain, Riskfuel’s clients avoid having to retrain their neural network to get accurate pricing under new market conditions. In addition to giving their customers more real-time answers and reducing compute spend, Riskfuel’s DNN models and system also helps in a few other ways, including:

  • Giving users the ability to compute Greeks faster. Greeks are a set of measures that show how sensitive the price of a derivative is to changes in different factors like volatility, time, interest rates
  • Finding problems in clients’ pricers by using the DNN to detect discontinuities in the price, as well as bugs in the model
  • Saving money with the agility to reflect changing conditions without wasting time or spending more money on compute
  • Facilitate portability of their software across platforms and operating system with minimal effort.

The Riskfuel DNN gives customers huge edges on their competition in pricing securities and exotic options better and faster. To get there, Chatigny and the Riskfuel team had to overcome some non-trivial machine learning challenges.

Tackling complex ML challenges

Riskfuel’s award-winning solution tackles a fundamental challenge in financial ML: ensuring neural networks can reliably price derivatives across complex, multidimensional domains. The team must navigate treacherous combinations of market conditions (like options with both high volatility and long maturity) where traditional pricing models often become unstable. For a bank’s trading desk, accuracy isn’t just desirable— it’s mission-critical across every possible market scenario. This requires sophisticated error analysis to understand not just where models fail, but why they fail.

Each Riskfuel client brings their own unique complexity: different data formats, testing methodologies, and deployment requirements. To manage this intricate web of models and metrics across multiple financial institutions, Riskfuel turned to Weights & Biases, using its experiment tracking platform to monitor model performance, identify error patterns, and ensure reliability across every corner of their pricing domains.

“W&B is the most robust and scalable way to track multiple machine learning projects, without having to handle a lot of infrastructure about tracking or tagging models,” said Chatigny. “When we have 20 plus+ projects and we want to be able to come back easily to figure out what kind of accuracy we got, what kind of behavior certain models produced in training, what was the code or model architecture used for training, – Weights & Biases has been super helpful for all that.”

Managing a complex workflow with Weights & Biases

At Riskfuel, Philippe Chatigny likens their organizational structure to a Formula 1 racing team, with three specialized groups working in precise coordination:
“Think of it like F1 racing,” explains Chatigny. “The quants are our drivers, the engineering team is our mechanics, and DevOps builds the race tracks.”

The quants, like F1 drivers, navigate complex financial terrain. These domain experts work directly with client pricing engines, mapping out mathematical domains, identifying model boundaries, and fine-tuning neural networks to achieve peak performance. They’re responsible for the crucial decisions that determine model accuracy and reliability.

The engineering team, which Chatigny is part of, serves as the mechanics, continuously optimizing the machine learning pipeline. They tackle research and development challenges across the entire workflow,w – from sophisticated data generation systems to making neural networks production-ready. Their work includes developing training pipelines and creating diagnostic tools to spot potential model failures before they impact trading decisions.

DevOps completes the trio by “building the race tracks,”” – constructing and maintaining the high-performance infrastructure that powers everything else. They manage the complex network of GPU clusters, training environments, and deployment systems that enable rapid model iteration and testing. This includes orchestrating the massive data generation needed for training and ensuring seamless model deployment to production environments.

Riskfuel’s machine learning pipeline reflects the complexity of the problem it solves, orchestrating multiple steps from initial client engagement to production deployment. The workflow begins by understanding how a specific client represents trades: – their data formats, pricing conventions, and risk parameters.

From there, the team builds and validates neural network models that can not only replicate the client’s pricing model but also serve results at the speed and scale required for real-time trading. This end-to-end process requires careful coordination between quants who understand financial mathematics, engineers who build the ML models, and DevOps teams who ensure infrastructure can handle production workloads.

The whole workflow consists of looking at different trades, checking the types of features that should be considered within that pricer, and then carving a domain that will encompass any financial market conditions. Then the process of generating data that targets that domain, training the neural network, and validating it. Once the model is validated, it is then packaged and deployed to different formats giving our clients a lot of flexibility to serve the model in any platform of their choosing.

“It’s a very iterative process between all these steps, and Weights & Biases helps with the training steps,” said Chatigny. “When we deploy training jobs on our Kubernetes cluster, W&B Models is really useful in helping us keep track of what model was run on the different projects that we’re working on.”

The team also relies heavily on the W&B Models dashboard, which serves as an important hub for all of Riskfuel’s ML activities.

“The workspace dashboard is the best feature for me,” said Chatigny. “It allows us to have a good dashboard of our model training and monitor it on a daily basis. To be able to filter our training runs, have the ability to add plots while training so we can compare runs on the fly, that’s really useful.”

Weights & Biases plays a key role in helping Riskfuel manage their complex multidimensional workflow, which in turn allows Riskfuel to focus more on their core mission, something they all take to heart. 

“We’re happy to win awards, it’s cool, but we’re really focused on solving that core problem, helping our clients and making a big environmental impact,” said Chatigny. “There’s an environmental personal objective that gets us motivated, and we’re very interested in solving the biggest problem facing a lot of mathematical functions used in the financial sector.”