Driving consistency and precision in fair value measurement with machine learning
As the financial landscape continues to evolve, the principles underpinning effective valuation practices remain just as critical—especially when navigating the complexities of illiquid financial instruments. While the surge of AI technologies has transformed many industries, the core concepts of leveraging machine learning for fair value measurement remain highly relevant. In this piece, we revisit our approach to using AI-driven solutions like hierarchical clustering and artificial neural networks to enhance accuracy and efficiency in valuing financial instruments without active markets.
The concept of fair value is one of the core principles of the International Financial Reporting Standards (IFRS), which guide thousands of companies around the world in preparing their financial statements. Unlike traditional historical cost accounting, which implies valuing assets at their price at acquisition and applying corresponding amortisation patterns to it, the fair value accounting attempts to reflect the current course of company’s business as accurately as possible, depending on the available pricing data. The fair value measurement is applied in various domains of the businesses’ operations, underpinning the logic of IFRS and improving the understanding of business performance by all the stakeholders involved.
The recognition of financial instruments is one of the accounting domains where fair value measurement gains particular prominence – both due to the contractual nature of these instruments and the nature of the markets they are traded on. According to IFRS 9, financial instruments are to be measured at fair value according to the fair value hierarchy, which for illiquid instruments suggests the use of price quotations for identical instruments for valuation purposes. In the absence of quotes on an active market, the standard prescribes turning to observable inputs such as quotes of sufficiently similar instruments. If there are no suitable assets or liabilities that could be used as a benchmark for valuing an instrument, unobservable inputs could be used to model the value of the instrument.
In this blog, we address the issue of measuring the fair value of instruments without an active market, but which have similar instruments quoted at active markets (Level 2). First, we use hierarchical clustering to find groups of similar instruments. As a second step, we train an artificial neural network (ANN) on the groups of similar instruments.
What is hierarchical clustering and how we apply it
Clustering is the art of finding patterns in data. Its goal is to:
- put similar objects in the same group, and
- put dissimilar objects in different groups
This notion is inspired by how we divide and categorise the world to make sense of it. For example, we would classify bicycles and cars as different means of transportation. At the same time, we could choose not to distinguish between them and categorise both as vehicles. A clustering algorithm mimics this process and tries to find meaningful categorisations for specific applications.
As mentioned previously, in case of absence of an active market for an instrument, the IFRS prescribes to look for similar assets or liabilities that are actively traded. However, how does one know which assets are similar? Also, how does a person know how many clusters of
similar instruments are there in the markets? To answer these questions, we built an AI algorithm comprised of two machine learning models, a Variational Autoencoder (VAE) and a Gaussian Mixture Model (GMM).
The VAE is the eyes of the AI. It can see high dimensional objects and compress them into a comprehensible, low dimensional space. By tuning how sharp it sees, we can make it see blurry enough to mistake similar objects to be the same, but clear enough to distinguish different objects. This is precisely what we want since cluster analysis is used to extract patterns in data. If the VAE is the eyes of the AI, then the GMM is the brain. Just like the human brain, the GMM struggles to comprehend high dimensional objects, but with help from the VAE, it can understand them. The GMM decides what objects to cluster and how many clusters there are by fitting multiple Gaussian distributions to the data and utilising the likelihood framework. In
short, we can think of this process as the VAE transforming the objects into dots on a paper and the GMM then colouring them. Compared to similar clustering algorithms, the VAE has a sharper vision, and the GMM is a smarter brain. To illustrate this, the algorithm is tested on the bond markets. Here, it manages to cluster the market and give information on whether to further merge or divide clusters. Figure 1 shows an example of a two-dimensional latent space for 19 clusters of bonds.
It also tells us which bonds are most representative of the cluster and which are outliers. Standard clustering algorithms do not do any of this by themselves. Figure 2 shows a cluster of bonds (the brown cluster in Figure 1) made by the algorithm.
After grouping similar bonds using hierarchical clustering, the idea is to train the ANN on a cluster containing similar level 1 bonds (to a level 2 bond in need of a benchmark) with a quoted yield to maturity of a bond of interest.
The input information would contain the available maturity, coupon and any other available bond features other than price and yield to maturity. The data would be processed by the trained ANN to estimate the resulting yield to maturity, which could be then used to calculate the fair value of the bond.
Such a procedure has several advantages compared to other methods. For instance, the pricing error is significantly reduced, and the model is expected to perform well with smaller data sets. The method can be further improved by using more “orthogonal” datasets (Natural Language Processing-based, or other alternative data).
Also, to ensure the acceptable performance of a solution, it is essential to create a rigorous model validation framework. We would suggest applying a simple linear regression as a model validation tool. It would benefit the solution in three distinct ways. First, it would be transparent: the model would be understood in terms of the regressors, which makes it easier to grasp which relationships are deemed relevant for the solution to work. Second, the interpretability of this model would be rather high – it could be explained in terms of the meaning and relative importance of each regressor. Finally, a robust and straightforward model such as ordinary least squares regression is an excellent tool for building trust with model users.
Read the original article here.
