Global Valuation is the first full platform implementation of Tensor Algebras for Financial Services
Global Valuation delivers for risk and pricing what Machine Learning has done for statistics. Esther takes the backward-looking AI / ML approach and applies it to future-looking risk and pricing.
Global Valuation will likely change the face of quantitative mathematics forever. Computing architecture has developed processors and memory tuned for Machine-Learning that are also tuned for the risk and pricing solver in Global Valuation. This delivers astonishing performance and the flexibility to easily solve a range of currently intractable use cases.
Global Valuation has defined EDSL, a modelling language for all use cases in risk and pricing which factors out all recursions, model solving and orchestration. This leaves the user to focus on their own business logic, including model definitions, payoffs and simulation parameters.
Our Vectorising Compiler massively outperforms multithreading rivals whilst also delivering flexibility to deploy onto a wide range of platforms. Just-In-Time compilation also enables enhancements in obfuscation and encryption of data passing between the Client data and model environments and the Global Valuation universal solver - opening up options for cloud execution in an increasingly regulated environment.
Global Valuation is designed to be the basis of a multi-billion $ platform that will disrupt Financial Services forever
Financial Services have accepted compromised shortcut pricing models for decades, since they could not marshal the compute power to solve the models fully. Global Valuation’s universal solver can look at all scenarios and all aggregations. Shortcut models will soon be a thing of the past, since risk and pricing compromise is unnecessary with rapid calculation of over a billion paths.
By using the same matrix multiplication as AI, but looking into the future, Global Valuation delivers orders of magnitude increase in scenario capability, mathematical stability and visualisation combined with significant hardware efficiency and savings in compute time.
Current quant platforms use forward induction methods, such as MonteCarlo simulations, to determine the potential future values of a portfolio. By sampling over 10,000 possible paths and using Probability Density Equations, reasonable approximations of the future state can be achieved. But these calculations are only marginally stable and require the manipulation of huge amounts of data over and over again to compute each path.
Implementation of these quant models requires huge computing grids, yet the problem is not compute bound, since it is the data manipulation to loop through each path that is the limiting factor. Shortcut models (Black Scholes, Heston, etc) have reduced the linear processing loops to explore each path, but current approaches are laborious and compromised.
Global Valuation takes a different approach
Before embarking on MonteCarlo simulations, Esther computes each matrix that represents all possible values in the future state at all relevant time points. These solution matrices represent each risk factor (Interest Rate, Yen, $, etc) required for the portfolio at each time point on the paths to the future.
These matrices are calculated using a strongly stable variant of backward induction and represent definitive and reusable tables for each risk factor to support nested MonteCarlo simulations. The tables are stored in memory and capture all useful forward information implied by calibrated market data and the codified portfolio of transactions. Further efficiency is achieved by sharing and recycling tables as much as possible throughout the portfolio, compressing and optimising cache strategies to streamline memory access as much as possible.
Generating Trade Pay-Offs, Calibrated Market Data and Probability Trees up front supports blisteringly fast traversal of nested MonteCarlo simulations using Tensor Algebra
By generating and storing the solution matrices, Esther can define the potential future values rapidly and accurately over a billion scenarios
Predicting and quantifying the potential paths of the future is much easier if you already know the solutions at each node.
Esther modules generate calibrated market data, trade pay-offs for the full portfolio and the probability tree as part of the pre-processing. This data can be re-used and added to for incremental calculation.
Forward induction is orchestrated using DAGs (probability trees) based on:
Risk factors relevant to the financial products within the portfolio, represented as solution matrices
Timepoints and optionality defined across all of the economic contracts in the portfolio
Forward cash flows defined from the trades
Different solution matrices are multiplied together for complex financial products with multiple risk factors, Esther performs forward induction MonteCarlo simulation to the timepoint for the first economic event using around 10,000 paths. At each of these nodes, 1000s of further nested MonteCarlo simulations will be performed.
This is repeated until all potential prices for the portfolio are determined, typically leading to over a billion scenarios. The potential paths are defined and traversed using DAGs to optimise processing time and ensure no duplications or gaps.
Clearly, this involves matrix multiplication on an enormous scale, leveraging modern compute capability and using fast exponentiation to minimise the number of mathematical operations. Since the solution matrices (previously calculated) are held in memory, the processing of over a billion paths can be performed in a fraction of the time that current platforms would require for 10,000 paths.
Pre-processed data and the solution matrices are held in memory. Esther challenges the compute to traverse the nested MonteCarlo probability tree and calculate the valuations at over a billion nodes at incredible speeds.