Machine learning in equity markets: what it can and can't do

What machine learning is good at in equity markets

Machine learning is a set of methods that learn patterns from data rather than following rules a person has written line by line. In equity markets, that ability is valuable for a simple reason. There's far more information than any person can process.

At iQuant.fund, which I co-founded and chaired, our research covered more than 20 exchanges and over 5,000 equities. No team of analysts could watch that universe continuously. A well-built model can.

In broad terms, machine learning helps with:

• Scale. Scanning thousands of securities across markets and time zones every day.
• Non-linear relationships. Many market relationships aren't straight lines. Some signals only matter in certain conditions or in combination with others. Methods such as gradient-boosted trees and neural networks can capture interactions that simple linear models miss.
• Text and other unstructured data. Natural language processing can read company announcements, earnings call transcripts and news far faster than a person.
• Consistency. A model applies the same rules on a volatile day as on a quiet one. It doesn't panic, and it doesn't get overconfident after a good run.

What it can't do is predict the future with certainty. Financial markets are noisy, and they change as participants learn and adapt to each other. The rest of this article is about working within those limits.

Momentum signals: an old idea with new tools

Momentum is the tendency for stocks that have risen over recent months to keep outperforming for a while, and for recent losers to keep lagging. It's one of the most thoroughly documented patterns in finance.

In 1993, Narasimhan Jegadeesh and Sheridan Titman published an influential study showing that buying past winners and selling past losers, measured over periods of three to twelve months, produced positive returns on average in US stocks. Later research found similar effects across many countries and asset classes. A common version measures returns over the past 12 months while skipping the most recent month, because very short-term moves tend to partly reverse.

Researchers usually distinguish between two types:

• Cross-sectional momentum ranks stocks against each other and favours the relative winners.
• Time-series momentum looks at whether each asset's own price trend is rising or falling.

Where machine learning adds value

Classic momentum strategies use one or two fixed lookback windows. Machine learning can test many windows together, combine momentum with other signals and adjust how much weight momentum receives in different market conditions.

Where momentum fails

Momentum has a well-known weakness: it can crash. When markets rebound sharply after a steep fall, the previous losers often jump the most. A momentum portfolio that is short those losers can suffer heavy losses in a very short time. Kent Daniel and Tobias Moskowitz studied these momentum crashes, including the one that followed the 2009 market bottom.

A momentum model that doesn't account for this risk is incomplete, however good its backtest looks.

Market regime detection and why it's harder than it looks

Markets move through different states, often called regimes. A low-volatility, steadily rising market behaves very differently from a high-volatility sell-off. Correlations between stocks shift, the signals that work shift, and so does the sensible level of risk.

Regime detection tries to identify the current state from data such as volatility, correlations, credit spreads and market breadth. Common statistical approaches include hidden Markov models, which assume the market switches between a small number of hidden states, and clustering methods that group similar periods together.

The idea is sound. In practice, several problems get in the way.

• Lag. A regime is often only clear in hindsight. By the time a model is confident a new regime has begun, much of the move may be over.
• Rare events. Crises don't happen often, so there are few examples to learn from, and each one tends to differ from the last.
• False alarms. A model tuned to catch every shift will flag many that come to nothing, and changing positions too often costs money.

My view is that regime analysis works best as context rather than as a trigger. It helps a portfolio manager understand what kind of market they're in and how much to trust other signals. Treating it as an alarm that says exactly when to act asks more of it than it can deliver.

Overfitting: the biggest risk in machine learning trading

If I had to name the single biggest danger in applying machine learning to markets, it would be overfitting.

Overfitting happens when a model learns the noise in historical data instead of a genuine pattern. It looks brilliant in a backtest and disappoints in live trading. Financial data is especially prone to it, because the signal-to-noise ratio in returns is very low and the underlying relationships change over time.

The problem grows with every test. Try enough strategies on the same historical data and some will look excellent purely by chance. Researchers including David Bailey and Marcos López de Prado have shown how quickly the odds of a false discovery rise as the number of trials grows. They proposed adjustments, such as the deflated Sharpe ratio, to account for it.

How serious teams guard against overfitting

• Out-of-sample testing. Hold back data the model never sees during development, and test on it once at the end.
• Walk-forward testing. Train on one period, test on the next, then roll forward. This mirrors how a model is actually used.
• Purged cross-validation. Standard cross-validation can leak information in time-series data. Purging and embargo methods remove overlapping observations between training and test sets.
• Economic reasoning. A signal with a plausible explanation, such as a behavioural bias or a structural constraint, is more likely to persist than one found purely by data mining.
• Counting the trials. Record how many variations were tested, and judge the final result with that number in mind.
• Simplicity. When two models perform similarly, the simpler one is usually the safer choice.

Data quality problems that break models

A model is only as good as the data it learns from. Market data looks clean on a screen, but it's full of traps.

Survivorship bias

If a historical database only contains companies that still exist, it has quietly removed every company that went bankrupt or was delisted. Backtests run on that data look better than reality. Good research databases include delisted securities.

Look-ahead bias

A backtest must only use information that was available at the time. Company financials are published weeks after a quarter ends and are sometimes restated later. Using the final, restated figures as if they were known on the last day of the quarter gives a model information it couldn't have had. Point-in-time data, which records what was known on each date, avoids this.

Corporate actions and exchange differences

Share splits, consolidations, dividends, mergers, spin-offs and ticker changes all need careful handling, or price histories will show false jumps. Across 20 or more exchanges, a model also has to deal with different currencies, trading hours, public holidays and time zones. A signal calculated with a stale price from a market that has already closed is a quiet but common source of error.

Costs and liquidity

A backtest that ignores brokerage, bid-ask spreads and market impact isn't a real test. Many strategies that look profitable on paper fade once realistic costs are included, especially in smaller, less liquid stocks where large orders move the price.

Why human oversight matters and why this belongs with institutions

Machine learning doesn't remove the need for human judgement. It changes where that judgement is applied.

People decide what a model is trying to achieve, which data it can use and what risks are acceptable. People check whether an unusual output is an insight or an error. People also decide when conditions have changed so much that a model's history no longer applies, and they remain accountable to clients and regulators when things go wrong.

Regulators take this seriously. The US Federal Reserve's 2011 supervisory guidance on model risk management, known as SR 11-7, sets out expectations for independent model validation, ongoing monitoring and clear governance. Its principles apply well beyond banking.

This is why I believe AI trading technology belongs with institutions. iQuant served institutional clients only: family offices, allocators, funds and proprietary trading teams. These organisations have investment committees, risk managers and compliance teams, along with the expertise to question a model rather than simply follow it. They can combine quantitative signals with their own research, size positions responsibly and plan for the fact that any model will sometimes be wrong.

Individual investors are in a different position. A signal presented without context about its failure rate, costs and risks can encourage exactly the wrong behaviour. Handing sophisticated models to people who can't evaluate them mostly shifts risk onto those least able to carry it.

Used well, machine learning makes good research teams faster and more thorough. It works best with experienced people setting its direction and checking its output. For more on how iQuant came about, read inside iQuant.fund.

About the author

Adrian Campbell is an Australian entrepreneur based in Indonesia and the founder and CEO of Kinnara, a global property marketplace. He co-founded iQuant.fund, which built AI-assisted systematic equity research for institutional clients, and served as its Chairman. Read his biography.

This article is general information only and is not financial advice.