Is AI the latest ‘holy grail’ in investing?

OVER the past four decades, I have seen investors chase successive “holy grails”, from portfolio insurance in the 1980s and risk parity in the 1990s, to smart beta in the 2000s.

None repealed the fundamental trade-off between risk and return.

Today, artificial intelligence-driven investing is touted as the next holy grail. It would be easy to dismiss this as another turn of the same cycle, but that would be premature.

AI can analyse information at a scale and speed previous investment tools could not, and dismissing it may prove more costly than embracing it too quickly.

Machine-learning models have outperformed traditional regression techniques in several prominent academic studies of return prediction.

Yet, much of the evidence comes from historical data, and whether back-tested success translates into durable real-world performance remains contested.

AI is likely to become more widely used in investing. But as its use spreads, it may improve the investment process without delivering a durable edge.

The problem is not whether AI can find predictive signals. It is whether those signals remain valuable once others can find them too.

In a study forthcoming in the October issue of the Journal of Financial Economics, Alejandro Lopez-Lira and Tang Yuehua found that strategy returns associated with large-language-model signals declined as adoption rose.

The pattern is consistent with – though does not prove – faster incorporation of information into prices. It suggests that even a genuine signal can lose economic value, as comparable tools become more widely used.

Rising adoption can dull the edge

That is not merely an implementation problem. It reflects a fundamental feature of markets: Once an advantage becomes widely known, competition erodes it.

History offers a useful parallel. In a 2016 Journal of Finance study, R David McLean and Jeffrey Pontiff examined 97 published predictors of cross-sectional stock returns.

Portfolio returns were 26 per cent lower out of sample and 58 per cent lower after publication. After allowing for the out-of-sample decline, the authors attributed an estimated further 32 per cent reduction to publication-informed trading.

Their study did not examine AI-generated signals, but it illustrates how an investment edge can weaken once it becomes widely known and exploited.

Scale, proprietary data and implementation expertise can still confer an advantage. That is the strongest case for the large and sophisticated, quantitative-driven institutions.

But most other investors would rather buy than build their AI capabilities, often drawing on the same models, research and vendors. Tools that are widely available may improve efficiency without creating enduring differentiation.

AI therefore presents investors with two distinct risks: A genuine signal may be competed away, while an apparent signal may never have been genuine at all. That is where backtesting has to be approached with caution.

In a 2014 article for the Notices of the American Mathematical Society, David Bailey, Jonathan Borwein, Marcos Lopez de Prado and Zhu Qiji Jim warned that testing enough strategy configurations can produce a strong simulated performance, even without a robust underlying edge.

Because analysts seldom disclose how many configurations they tried, investors may be unable to judge the extent of selection bias. The authors further showed that, under memory effects in financial time series, an overfitted strategy can have negative expected out-of-sample returns.

AI greatly expands the number of strategy variations researchers can test. Without equally rigorous validation, that capacity can materially increase the risk of overfitting.

Ask how many versions were tested, before this one was selected.

Then, ask whether the researchers used genuinely untouched test data, allowed for trading costs and tested the strategy in different periods and market conditions.

The role of human judgment

Yet, forecasting is only one part of investing.

Human judgment remains essential in translating forecasts into position sizing and liquidity considerations in a portfolio, including risk limits, and in taking responsibility for those decisions.

The job is to commit money against liabilities that fall due whether or not a forecast proves right, on behalf of clients who may want their capital back at the worst possible moment.

Over time, results turn not only on the quality of a signal, but also how much you bet and how long you can hold the position without being forced to sell. It is not uncommon to have come across good analysis ruined by bad sizing.

Even an accurate forecast does not determine how much capital to commit, how much liquidity to retain, or how much loss an investor can withstand. Somebody must make those decisions and answer for them.

A model can help diagnose why a strategy failed, but it cannot bear responsibility for the judgment made or the losses incurred.

Deploying an AI-powered framework does not remove accountability, it makes accountability harder to exercise when decision-making is opaque or responsibility is diffused.

None of this means investors should not use AI. It means they should use it for what it is.

AI is among the most powerful tools I have encountered in my career. Its most dependable value may lie not in producing a permanent forecasting edge, but instead in strengthening research, monitoring risk and challenging investment assumptions.

It can process far more filings and transcripts than a human team could review unaided, monitor exposures continuously rather than quarterly, and give analysts back time otherwise spent gathering information.

Ask it to assess an investment case and, in my experience, it often finds the weak seam quickly, sometimes one the team that built the case can no longer see.

That is not a small prize. It is an enhanced research process, not a machine that prints money.

That is the test investors should apply.

If a manager says AI has changed how the company works, take the claim seriously. If the manager says it has changed what the business can earn, ask what makes that advantage scarce, durable and accountable.

Among investing’s supposed holy grails, AI may be the most powerful yet. But its lasting value will lie in improving investment decisions, not in eliminating judgment, risk or accountability.

The writer is chief investment officer, DBS Bank