ESG and macroeconomic financial models
Client situation
A Tier 1 leading AM in ESG and green investment would like to develop a clear AI-based methodology & tools in order to integrate ESG factors as a structuring element to anticipate profitability and manage medium and long-term risks.
They realize that they are dealing with ESG as an exogenous factor. They would like to understand the relationship between ESG factors and other quantitative driving factors (macroeconomic, microeconomic, etc.), and integrate their own qualitative knowledge in the model, in order to generate alpha.
Scalnyx solution
We customized our Projector platform based on their AM team experts needs to match their database and also their portfolio and risk analysis frameworks.They used the platform to build Causal AI models from non-linear factors such as ESG, micro/macro-economic, and in-house portfolio data. They tested the complex combination of factors and also elaborate on the complex mixture of environmental and economic factors. They assess the accuracy of each model with a back-testing framework.
The generated Causal AI model inferences are fully explainable and interpretable.
Client gain
The resulting Causal AI models improved by the domain experts knowledge (asset manager), outperformed the performance of classical linear models and other AI/ML techniques, proving the interest of quantamental investing if the asset manager disposes of the right tools.
The company appropriates the construction of their own tools and methodologies that match the company’s culture, governance, and team organization, and also their data providers and sources.
Derivative pricing
Client situation
A medium-sized broker is using real time futures and options valuation for an arbitrage desk and financial engineering business. They dedicate a bunch of ultra fast servers to value derivative products based on heavy analytics methods using monte-carlo simulations.
They are facing several issues: infrastructure is too expensive; calculation times are too high and traders are complaining that some competitors are pricing faster.
Scalnyx solution
We used our ML algorithm acceleration toolbox to approximate option pricing with stochastic volatility models.
Based on the given training data, the toolkit generates a corresponding Deep Learning model solution that matches the underlying process; it estimates a function between the input data and premiums, minimizing a given cost function (usually the mean squared error between the model price and the observed price on the market) to reach a good out-of-sample performance.
The toolkit can even train a DL model with a small data set sample.
Client gain
End-to-end valuation and risk sensitivities computation is 800 times faster than numerical computations and accuracy is optimized. Hence, the bank can provide on demand real-time valuation to the desks within a few seconds instead of submitting requests to the system.
Each ML-based derivative pricing instance can run on 1 CPU core instead of several cores required for a complex monte-carlo simulation, hence reducing hardware footprint and optimizing operational efficiency.
Internal model FRTB
Client situation
In the context of the recent BASEL IV regulations, a bank requires the FRTB risk calculations of Non Modelisable Risk Factors (NMRF, which are typically illiquid assets). With no other options available at the moment, the client currently uses an approach which leads to an unnecessarily high-penalized capitalization: Capital requirements for each non-modellable risk factor (NMRF) are determined under stress scenario that is calibrated to be at least as prudent as a «pessimist» (expected shortfall) model and without any diversification to the internally modeled risks . Nonetheless, the BASEL IV standard allows the approach of risk models specifically tailored for specific assets, which could justify a more convenient capitalization.
Scalnyx solution
We adapted our ML algorithm acceleration toolbox to FRTB NMRF needs in order to allow the identification of proxy variables, such that the risk is computed through the proxy data and not through the original risk factor itself. For our particular case, ie. a NMRF, the objective is to identify sufficiently similar modelisable risk factors (eg. choosing among assets of the same type and region, similar in nature, but with a better liquidity). Since the chosen proxy variables are modelisable, we are allowed by the standard compute (up to a percentage) of the capitalization of a NMRF as a modelisable risk factor, which allows us to take advantage of diversification, and thus justifying to the regulators smaller FRTB penalizations.
Client gain
The resulting accurate proxies based on machine learning reduce the capital impact of NMRF. In addition, the bank fully controls the modeling process and generation of NMRF data for application of FRTB rules.
As a result, the bank gains in effective productivity and ownership of the full process, and reduces massively the liabilities for regulatory reporting.
Scalable matching engine
Client situation
A stock exchange is facing incompatible constraints for its trading infrastructure: growing number of orders and executions combined with increased complexity of operations, leading to an increasing need to invest in more and new machines. However, the technical architecture is not designed to scale and needs to be revisited.
In addition, growing competition requiring low latency guaranteed execution and cost-cutting is pushing to reduce hardware footprint and optimize software design.
Scalnyx solution
We used our Parallel computing SDK to provide the core technology for redesigning the matching engine technical architecture to enable low-latency and scalability based on reactive systems principles.
The developers easily built their matching engine business functionalities with our multicore-optimized SDK and runtime environment, while avoiding the complexity of multithreading development. The new architecture is by nature scalable and very safe, thanks to our highly simplified technical architecture, which massively reduces the number of servers needed.
Fast and fully predictable response times avoiding the complexity of multithreading is resulted, and also compatible with multi-core chip architectures.
Client gain
Fast and fully predictable response times avoiding the complexity of multithreading is resulted, and also compatible with multi-core chip architectures.
Hardware infrastructure is reduced and latency is no longer dependent on number or complexity of operations. The micro services architecture can be easily monitored and adapted in real time. The system is asynchronous by nature and self healing, thus an end to end latency is reduced and 100 % deterministic.
Software development is made easier, operational control is improved and over IT budget is reduced.