Last time out, we looked at what wealth and asset managers should think about when using artificial intelligence (AI) to construct and manage their equity books. As we have also covered, however, AI brings different characteristics to different asset classes, so let’s now consider its applicability to the world of fixed income.
With fixed income typically being more complex, AI’s application to this market must be viewed through a slightly different lens. Bond markets are directly shaped by multiple moving parts – macroeconomic factors such as interest rates and inflation, or credit quality at the country and company level. Bond prices are also influenced by market technicals such as liquidity conditions and supply-demand dynamics.
Liquidity is an important feature too. Some of the very large government bonds – of certain maturities, at least – are clearly very liquid. For much issuance across fixed income, however, liquidity is limited, typically resulting in wider bid-ask spreads and less transparent price discovery than in equity markets.
A swing of even a few basis points can have a significant impact. Perhaps this is an inherent feature of the market, with prices mostly determined by a combination of inflation, credit and term structure. In addition, a meaningful share of returns comes from carry and roll-down effects, as well as interest rate sensitivity – as measured by duration – which adds complexity to how portfolios behave.
The other defining feature of bond markets is that changes to the overarching macroeconomic or political environment can lead to rapid repricing in rates and credit spreads, even if overall volatility is often lower than in equities.
Enriched credit analysis
Where AI can help versus traditional modelling in fixed income is twofold. It can capture and adapt to sudden market shifts that may be missed or lagged by traditional models; and it can enrich credit analysis by incorporating a wider set of data.
As an example, traditional fixed income models – often based on relatively few factors such as duration, credit spreads or carry – might only capture changes brought on by volatility. Yet an AI model can adapt more quickly to emerging patterns in the data, such as policy changes or inflation shocks.
The other area it can be beneficial is around credit analysis. Historically we have priced credit based on the likelihood of a company defaulting or not. Yet the problem with that is one of opacity – bonds with higher credit risk are typically less liquid and therefore you have fewer observations on which to base your price.
“Of course, it cannot give you anything that does not exist – AI is not a crystal ball – but it can help extract more information from what is available.
It is useful to have a reliable toolkit when things get difficult – but do be ready to pause, modify or change direction altogether if the underlying assumptions or functioning of the models are impaired.”
With AI, you can now factor in other sources of relevant information, such as newsflow or non-financial company information. It is also helpful where you may have scarce or noisy data. This can, in more traditional linear models, present a false picture. This means less reliance on approximations and analogies and an improvement in the richness and overall accuracy of your models.
If a vendor looks set to go bankrupt, for example, valuations can vary widely. Recovery expectations, capital structure and market conditions or small changes in assumptions can have a large impact. Any improvement to accuracy will help investors make significantly better-informed decisions.
You may only have access to partial information. After default, bonds often trade in less transparent, dealer-driven or private markets, where price discovery can be fragmented. An AI model, however, can use this other information, like news or other variables, to help price with greater accuracy. Of course, it cannot give you anything that does not exist – AI is not a crystal ball – but it can help extract more information from what is available.
Two main protections
If concerns exist about mismatching certain signals, we see two main protections you can introduce to minimise the chance of any correlations being wrongly made. First, you need technical processes to avoid blindly mining data or capturing bias. This includes guarding against overfitting or spurious relationships. Second, you need to bring in human judgement to manage and govern the process appropriately.
A model might explain historical data well, for instance, but still fail in prediction. Experienced professionals who understand the models – including their limitations – and the market environment are needed to review process and outputs to blend and scale them with related ones. They can also then step in if things do not add up or exceptional market circumstances occur.
This is a constant consideration when using AI to support your processes – think of it as a useful tool to add to your mix, not an infallible resource. It is useful to have a reliable toolkit when things get difficult – but do be ready to pause, modify or change direction altogether if the underlying assumptions or functioning of the models are impaired.
As I alluded to at the start, since fixed income markets are heavily affected by the political landscape and macroeconomics – both outside of their control and over which their influence is limited – this adds complexity. Therefore, no single model ever can fully tell you what is going on.
Most advanced AI models capture reality better than traditional ones, but are harder to interpret. They will not give you an easily articulated ‘cause-and-effect’ but they might give you the factor that mattered the most. This might have been inflation. Or was it a Fed decision?
And it can tell you how those factors came together to influence a particular position or risk factor – and to what degree. Such a level of explainability is crucial for the human being brought into the fold who needs to understand how the models arrived at their results.
AI should therefore be seen as a necessary complementary tool – one that can broaden the information set available to investors and support better-informed decisions, rather than fully replace all existing approaches and techniques.
Giovanni Beliossi is head of investment strategies at Axyon AI
