Factor investing with AI means using machine-learning models to find, weight and rebalance the stock characteristics (value, momentum, quality and others) that have historically driven returns. The models can read more data than a person can and adapt faster. The evidence that they reliably beat the market after costs is mixed, and the risks are real. Below we explain the mechanics, the research, the pitfalls and a short checklist you can run on any AI-driven strategy, including ours.
TL;DR
- A factor is a measurable stock characteristic, such as cheapness (value), price trend (momentum) or profitability (quality), linked to differences in returns.
- AI adds three things: it reads unstructured text, it finds nonlinear relationships, and it can adjust factor weights as conditions change.
- Academic results are promising on paper. Studies of real AI-labeled funds found no benchmark-beating performance, so treat backtests as hypotheses.
- The main risks are overfitting, high turnover, opaque models and taxes in taxable accounts.
- Before you follow any AI strategy, check transparency, turnover, costs and what happens in a downturn.
Table of Contents
- What factor investing is
- Where AI fits in
- Which factors AI strategies analyze
- How an AI factor strategy picks stocks and rebalances
- What the evidence says about returns
- Risks and limitations to understand
- Costs and taxes
- A five-point check before you follow any AI factor strategy
- Our perspective
- Following a rules-based AI portfolio with StockPilot
- FAQ
- Sources
What factor investing is
Factor investing seeks exposure to the underlying forces that drive portfolio risk and returns. MSCI's research separates three uses: factor models that manage risk, factor strategies that try to capture factor returns, and factor allocation that uses factors to drive asset-allocation decisions.
In practice, most investors meet factors as a screen or a rule. Instead of asking whether one company is a good business, you ask which groups of stocks share a trait that has paid off over long periods. Then you hold a diversified basket with that trait.
The three factors you will see most often:
- Value: companies trading at low prices relative to fundamentals.
- Momentum: companies with strong market and analyst sentiment.
- Quality: companies with efficient management, solid profitability and strong cash flows.
That description comes from Northern Trust's 2025 factor research, which uses U.S. data from December 2005 to December 2023.
Where AI fits in
Classic factor investing uses a short list of hand-built signals, combined with simple weights. The weakness is that relationships change. The authors of Machine Learning for Factor Investing note that future returns may depend on firm characteristics whose link to performance is unknown and varies over time. Machine learning is useful for spotting hidden patterns and retraining as the data changes.
Three specific contributions stand out.
Nonlinear relationships
An ordinary model assumes that if cheaper is better, then twice as cheap is twice as good. Robeco's research says machine-learning models can beat linear ones because they capture nonlinear relationships and higher-order interactions among features. For example, momentum might work well in profitable companies and poorly in unprofitable ones.
Unstructured data
An asset manager's 2023 U.S. guide to AI in quantitative investing says AI can extract signals from structured financial data and from text: news, analyst reports and earnings-call transcripts. It adds that natural-language processing can pick up information in call transcripts that the official earnings release does not contain.
That is the idea behind earnings call narrative analysis, which scores what management says against what the company later delivers. We have written separately about why basic sentiment models fall short for investing.
Dynamic weighting
Instead of fixing weights once, a model can change them as new data arrives. This is called factor timing, and it is the hardest part to get right.
Which factors AI strategies analyze
Most AI-driven strategies start with the familiar factors and add many more inputs. A December 2025 presentation on an AI-enhanced value strategy describes inputs spanning fundamental, market and macro, sentiment, and ESG data across several time horizons. Its examples include liquidity, volatility, money flow, volume, momentum, reversal, macroeconomic indicators, relative fundamentals, cash flows, analyst expectations, revisions and sentiment.
The inputs that matter most are often plain. Robeco reports that short-term reversal, stock and sector momentum, return volatility and firm size are among the features that most consistently drive cross-sectional return predictions in its experience and the studies it cites. If you want to see how a varied input set gets assembled and tested, equity curve analysis methods show one way to judge whether a rule's results hold together over time.
| Factor | What it measures | Typical AI contribution |
|---|---|---|
| Value | Price relative to fundamentals | Blends many valuation measures and revisions |
| Momentum | Price trend and analyst sentiment | Detects when trends are fading |
| Quality | Profitability, management efficiency, cash flow | Reads text and filings for early signs of change |
| Volatility | Size of price swings | Adjusts exposure as risk shifts |
| Size | Company market value | Interacts with other factors |
How an AI factor strategy picks stocks and rebalances
Strip away the jargon and the process follows the same loop as any portfolio management. A 2024 literature review describes it: set an investment policy and constraints, analyze market and security risk and return, allocate capital and select securities, execute trades, evaluate performance, update market conditions, and rebalance. The review stresses that feedback and rebalancing are necessary, because the process checks whether the original policy still fits.
AI changes the analysis and selection steps, and in some cases the others too. The asset manager's guide above says machine learning can also control portfolio risk, lower execution costs and optimize positioning.
One design choice deserves attention: how much the model trusts itself. Northern Trust describes a regularization-based timing process that adjusts factor weights as data arrives. When the model is skeptical, it can reduce or turn off timing and revert toward equal factor weights. We like that principle: when the evidence weakens, the model shrinks toward a simple, diversified baseline instead of forcing an aggressive signal.
What the evidence says about returns
This is where claims need the most care. The research points in two directions.
The encouraging results
- Northern Trust reports that its hypothetical dynamic factor-timing portfolio beat the Russell 1000 excluding financials and an equal-weighted factor portfolio. It stresses that the results are illustrative, not actual performance.
- Morningstar reported in October 2025 that machine-learning factor-selection strategies earned 0.27% to 1.39% per month in high-minus-low factor spreads, depending on the algorithm. Factor momentum explained most of the predictability, and the strategies had to replace 37% to 66% of factors each month.
- A December 2022 NBER working paper using U.S. equities from 1981 to 2020 reports out-of-sample Sharpe ratios near 3 and monthly alpha around 0.80% for its tree-based models.
- A September 2023 NBER paper reports that models with extremely large numbers of parameters beat simpler alternatives by a wide margin in out-of-sample pricing.
These are research-grade results, mostly gross of trading costs, and several rely on turnover that a real investor would struggle to implement.
The sobering results
- A November 2024 working paper on U.S. mutual funds labeled as AI found they performed similarly to comparable human-managed funds and did not beat a market benchmark. The funds showed better market timing but worse stock selection.
- A September 2024 experiment using anonymized U.S. data found that Google Gemini 1.5 Flash did not significantly beat a naive portfolio or the S&P 500 on returns or Sharpe ratios, and performance declined as the horizon lengthened.
- Even the paper behind the large gains from public information makes the same point about frictions. Its authors, in a version revised August 29, 2026, found that counterfactual portfolios for 3,254 active mutual funds earned 2.7% per quarter unconstrained. After realistic style, risk, position-size and trading-cost constraints, attainable value fell to 0.86%.
A 2024 study in Quantitative Finance reported a Sharpe ratio of 4.49 for factors that ChatGPT conceptualized, according to a research index entry. We would not build a plan on a number like that. Sharpe ratios that high in a backtest usually signal that the test is easier than real life.
The fair reading: AI can improve how factors are selected and timed, but the gap between a paper result and a live, after-cost result is wide. Northern Trust itself warns of misapplication, overreliance on non-transparent estimates and overfitting.
Risks and limitations to understand
Overfitting
A model with enough knobs can explain the past perfectly and predict nothing. The CFA Institute's summary of machine-learning stock-selection research names two defenses. Forecast combinations aim to improve reliability when different algorithms and training sets find similar relationships. Feature engineering narrows the questions and improves the signal-to-noise ratio.
High turnover and opacity
Robeco's 2023 commentary warns that machine-learning factor models trained to predict next-month returns can trade excessively, and their outcomes may be hard to interpret. The asset manager's guide adds that AI-based investing requires transparency: you should be able to break down features, positions, risks and performance.
Model and data errors
A U.S. investor guide from October 2024 warns that quantitative models can miss unforeseen market factors or interactions. It also says model mistakes, data problems, software issues and programming errors can go undetected until they hurt performance. It lists further AI-specific concerns: privacy, bias, security and misuse.
Stress periods
The SEC's 2020 Report to Congress on algorithmic trading says algorithmic trading can improve liquidity, execution and risk management in normal U.S. markets. It also says some algorithmic trading can worsen unusual stress or volatility, and complex interconnected systems can fail operationally.
For a broader look at what can go wrong, see our guide to AI investing risks.
Costs and taxes
Turnover has a price. A fund prospectus filed with the SEC notes that fund shares may carry annual operating expenses and brokerage commissions, and that higher turnover may raise transaction costs and taxes in taxable accounts. Short-term gains are taxed at ordinary income rates.
In practice, a strategy that replaces a large share of its holdings each month may look strong before costs and ordinary before taxes. If you plan to follow one in a taxable account, ask how often positions change. If you can, use a tax-advantaged account. A tax professional can tell you what applies to your situation.
A five-point check before you follow any AI factor strategy
- Can you see the holdings? You should know every position, its size and why it is there.
- What is the live record? Prefer results from real trading over backtests, and look at drawdowns as well as returns. Our guide to portfolio analytics metrics that reveal skill shows which numbers to ask for.
- How often does it trade? Turnover drives costs and taxes.
- What does it do in a downturn? Look for a written rule, such as cutting equity exposure, not a promise.
- Who holds the money? You should keep custody and approve every trade.
Our perspective
We think the research supports a modest, practical conclusion. AI is useful for processing many kinds of data and for enforcing rules consistently. It is not a prediction machine, and the studies of live AI funds back that up.
That is why StockPilot is built around a fixed structure instead of an endless stream of picks. Our model portfolio holds 19 positions: six income-producing positions and 13 AI-selected growth positions. The system evaluates market, fundamental, technical, sentiment and statistical data, and it issues after-hours buy and sell alerts only when predefined investment and risk criteria call for a change. In unfavorable conditions it can reduce growth exposure and move capital into Treasury ETFs.
We care most about two things the research flags: transparency and restraint. You can see the whole portfolio, you place every trade in your own brokerage account, and a rule has to trigger before anything changes. If you want to see how we think about momentum inside that framework, read our piece on the 19-position AI momentum portfolio, and on how we manage concentration, our portfolio risk management guide.
— David Elliot
Following a rules-based AI portfolio with StockPilot
If you like the discipline of factor investing but do not want to build or monitor models yourself, StockPilot Premium gives you the model portfolio, position sizing guidance, performance analytics and after-hours alerts for $147 per month. You mirror the trades in your own brokerage account, and we never take custody.
See how it works on the StockPilot Premium page, and read the risk disclosure before you decide. Investing involves risk, including loss of principal.
FAQ
Is factor investing with AI better than ordinary factor investing?
Not necessarily. AI can capture nonlinear effects and adapt factor weights, which is a real advantage on paper. But live results are less clear: a 2024 working paper found U.S. AI-labeled funds did not beat a market benchmark. Judge any strategy by its after-cost, live record.
Can I do this myself with ChatGPT?
You can generate ideas, but research on language models as stock pickers is not encouraging. One 2024 experiment found Gemini 1.5 Flash did not significantly beat a naive portfolio or the S&P 500. You would also have to handle data quality, position sizing, rebalancing and risk rules yourself.
Which factors matter most?
Value, momentum and quality are the established ones. Research also points to short-term reversal, volatility and size as consistent predictors. No single factor works in every period, which is why diversification across factors is common.
Do AI factor strategies cost more in taxes?
They can. Higher turnover may raise transaction costs and taxes in taxable accounts, and short-term gains are taxed at ordinary income rates. Ask how often the strategy trades and talk to a tax professional about your account type.
Sources
- MSCI: The Future of Factor Investing
- Northern Trust: Exploiting the Benefits of AI for Factor Investors
- Machine Learning for Factor Investing, introduction
- Robeco: Machine learning and data science applications in investments
- Pictet Asset Management: Quant AI portfolio construction
- WisdomTree: AI Enhanced Value Strategies presentation
- Frontiers in Artificial Intelligence: Enhancing portfolio management using AI
- Morningstar: Machine learning can forecast which stock factors will outperform
- NBER working paper w30805
- NBER working paper w31689
- SSRN: Assessing the Performance of AI-Labelled Portfolios
- SSRN: Can AI Beat a Naive Portfolio?
- SSRN: The Unexploited Value of Public Information
- RePEc: GPT's idea of stock factors
- CFA Institute: Machine Learning for Stock Selection summary
- Robeco: Why the best is yet to come for factor investors
- Voya: A Guide to Investing with Artificial Intelligence
- SEC: Report to Congress on Algorithmic Trading
- SEC filing: fund prospectus on expenses and turnover