Trades Fail: Tackling the Problem with AI Operationalized
Trade failures remain a chronic problem in financial operations. With the new T+1 mandate, the problem is likely to become even more acute. Our goal with this PEP solution is to help mitigate this issue. For us, this means marrying our deep understanding of settlement trade workflows with the right AI/ML tools to benefit the operations teams charged with ensuring trade settlement happens smoothly.
The Last Mile Aspect
Securities trades fail for many reasons. Historically, trade failures are not predicted but rather addressed only after they fail to settle. Furthermore, those charged with fixing fails must begin their work by acting as detectives to determine the root cause of the trade failure. We knew that one driver of fails was bad SSIs and hypothesized we would add value to our customers if we could (1) predict when a bad SSI would cause a trade to fail and (2) provide the likely root cause of the bad SSI.
Starting the Solutioning Journey
Predicting whether an SSI is “good” or “bad” is a probabilistic classification problem. Solving the classification problem required both (1) selection of a classification model type (e.g., tree-based, deep learning, etc.) and (2) input features that the model would use to make its predictions. Beyond the classification model, we also needed an orchestration layer that would integrate the model into customer workflows as part of a broader end-to-end, last-mile solution.
Choosing the Right Models
Our development efforts began with the creation of synthetic trade and SSI data and have continued with actual trade data from a Tier 1 global investment bank Proof Of Concept (POC) partner.
With this data, we have trained and evaluated various models and methods, including decision trees, deep learning models, and Bayesian methods. The approach that produces the best results for predicting both “good” vs. “bad” SSIs and likely fail reason codes is an ensemble of separately optimized models.
Handling Categorical Features with High Unique Value Counts
An SSI can be broken down into a number of distinct component features (e.g. client account at custodian, etc.). These features present two key challenges: (1) they are categorical (i.e. essentially strings) and (2) have large numbers of unique values. While the obvious way to encode the categorical feature values is via one-hot-encoding, the large number of unique feature values makes this approach impractical (i.e., would lead to an unstable model). Accordingly, the data scientists at CTFSI have developed more advanced and creative techniques to represent SSIs as numeric inputs to our models. Additionally, the CTFSI team has created a number of engineered SSI features from the historical trading data to improve the model performance.
Looking Ahead
The journey to accurately predict the likelihood of, and the reason behind, SSI-driven trade fails has been complex and challenging. However, using CTFSI’s unique approach – what we call “AI Operationalized” – we have been able to design a solution that is highly pragmatic and actionable. We are taking the same approach – building on our Predictive Exception Processing (PEP) platform – to tackle other causes of trade fails – including affirmations and inventory issues – along with many other exception processes across the middle and back office in financial services.
Stay tuned – there is much more to come!