Testing AI Stock Screeners: Do They Beat a Basic Value Filter?

By Chanuka Geekiyanage Oct 19, 2026

AI stock screeners claim to find undervalued stocks better than a plain filter built on price-to-book, earnings multiples, free cash flow, and debt. A basic value filter costs almost nothing and every rule can be audited, so an AI tool has to earn its place with better results after costs. The main dangers are backtests that use information the model could not have had at the time, short track records, and higher turnover. This article reviews the published evidence and gives you a way to test any screener against a simple value filter before committing money.

Panaprium is independent and reader supported. If you buy something through our link, we may earn a commission. If you can, please support us on a monthly basis. It takes less than a minute to set up, and you will be making a big impact every single month. Thank you!

What the Evidence Supports

No AI screener I found has shown a durable, independently verified edge over a simple value filter after costs. The strongest pro-AI evidence is narrow: language models can predict the direction of earnings from financial statements. That is a useful research aid, not proof of a better stock picker.

Treat the basic value filter as the benchmark every AI tool must beat. Judge any screener on these points:

Return after costs: include fund fees, trading costs, and taxes from turnover.
Drawdown and volatility: a higher return with a much deeper drawdown is not a clean win.
Out-of-sample results: performance on data not used to design or tune the rules.
Comparison to something ownable: an index fund or a plain value ETF you could have bought instead.

Win rate alone misleads, because a tool can win most picks and still lose money if the losers are larger.

The Contenders

The Basic Value Filter

An example filter might require a low price-to-book ratio, a price-to-earnings ratio below the sector median, positive free cash flow, and moderate debt. You can rebuild it in any screener and see why each stock passed.

Its weakness is that value is cyclical. Value stocks outperformed from 2001 to 2008, but recent years favored growth. In 2025, S&P 500 Growth returned 19.9% against 12.3% for S&P 500 Value. Over ten years to August 31, 2025, the Russell 1000 Growth Index returned about 17.9% a year against about 14.3% for the broader Russell 1000.

Image source: Hartford Funds

A filter also has a known trap. A stock can look cheap because the business is deteriorating, so add profitability and balance-sheet checks.

AI-Managed Fund: AIEQ

AIEQ is the clearest real-world test of an AI stock picker. It launched in October 2017, scans filings, news, and market data, and holds roughly 30 to 70 stocks.

The early record looked strong. After one year it was up 11.81%, just ahead of the S&P 500, and ten of its top fifteen contributors were too small for the S&P 500. Small-cap picks raise return and risk together, as covered in this piece on small-cap AI stocks, risk and opportunity.

The lead did not last. By May 2021, AIEQ had returned about 64% since inception against 75% for the S&P 500 Total Return Index. A Barchart article also listed an expense ratio of 0.75%, so any edge must cover that fee every year. Check the current fee and performance before relying on these older figures.

Language-Model Stock Analysis

The best academic case for AI is a University of Chicago study. Researchers gave GPT-4 anonymized financial statements and asked it to predict the direction of future earnings. With chain-of-thought prompting, it reached 60.35% accuracy, against 52.71% for analysts, while a simple prompt reached 52.33%.

Read this carefully. Predicting earnings direction is not the same as beating the market, because prices already reflect expected earnings. The plain prompt also only matched analysts, so results depend heavily on how the model is prompted.

Comparison

Option	Evidence	Main weakness	Best for
Basic value filter	Long factor history, but cyclical	Value traps, long stretches of underperformance	A transparent, low-cost benchmark
AI-managed fund (AIEQ)	Trailed the S&P 500 by May 2021	Opaque process, 0.75% reported fee	Small, experimental positions
LLM financial analysis	Beat analysts on earnings direction in one study	Earnings direction is not stock returns	Research support, not stock picking

Why AI Backtests Look Better Than They Are

The biggest problem is look-ahead bias. A model trained in 2024 may already “know” how 2019 stocks moved, so testing it on 2019 data proves little. One 2025 paper found that a one-standard-deviation rise in LLM prediction predicted a 0.197% higher next-day return, and memorization added an effect equal to about 37% of that baseline.

The fix is to test only on data after the model’s training cutoff, or use point-in-time data with masked company names. If a vendor cannot explain how its backtest avoided this, assume it did not.

How to Run Your Own Test

Write the scorecard before you see results.

Freeze the rules: fix the value filter and the AI settings or prompt, and do not change them after seeing results.
Use the same universe and schedule: same stock list, same monthly or quarterly rebalancing.
Test after the cutoff: use only periods after the AI model’s training cutoff.
Include costs: subtract fees and estimated trading costs for both.
Run it forward: paper trade through at least one meaningful market swing before using real money.

Situation	Recommended option	Why
You want an auditable baseline.	Basic value filter	Rules are visible and nearly free
You read many filings and need speed	LLM as a research assistant	Evidence supports earnings analysis, not stock picking
You want AI exposure without work.	Small position in an AI fund after checking fees and results	Evidence of an edge is weak
A vendor shows only a backtest	Do not buy yet	Look-ahead bias is likely unaddressed

Risks and Tradeoffs

An AI screener can drift toward the same mega-cap names as the index, so you may pay extra for something close to what you already own. That matters because the largest S&P 500 weights carry heavy AI exposure, as covered in this look at AI stock concentration risk in the S&P 500. Opaque models also make it hard to know why you own a stock, and frequent rebalancing adds fees and taxes.

My Take

Start with the basic value filter and use it as your benchmark. It is cheap, transparent, and has a long factor record, even though growth has led recently. Use AI only where it has documented value, such as reading filings or challenging your own thesis.

I would not pay for an AI stock-picking product on a backtest alone, and I would not hold an AI fund as a core position given AIEQ’s record and fee. Before using any screener, I would check for live out-of-sample results, point-in-time data, turnover, and how it compares with a plain index fund. If it passes and still beats your value filter after costs, it deserves a small allocation.

Conclusion

A basic value filter remains the right benchmark. AI tools have not shown a verified, lasting edge over it, and AIEQ’s record against the S&P 500 is a sober example. The best evidence for AI supports it as a research assistant, not a replacement for a rules-based filter.

The key risks are look-ahead bias, short track records, and fees and turnover. Your next step is to write down a simple value filter, freeze it, and run any AI screener against it on post-cutoff data with costs included.

FAQs

1. Can I trust an AI screener’s backtest?

Only if it uses point-in-time data and tests on periods after the model’s training cutoff. Without that, the results may reflect memorized outcomes instead of skill.

2. How long should I test an AI screener before using real money?

Paper trade it through at least one meaningful market swing if you can, since a few good months prove very little. Compare it with your value filter over the same period, with costs included.

3. Does a basic value filter still work after years of growth leadership?

Value has trailed growth over the past decade, including in 2025, but leadership has reversed before, as it did from 2001 to 2008. Treat it as a cyclical strategy that needs patience and quality checks.

4. Are AI-managed ETFs worth the fee?

AIEQ trailed the S&P 500 by May 2021 and has been reported with a 0.75% expense ratio. Check current performance and fees, and compare it with a low-cost index fund.

5. What should I do with an AI screener’s output?

Use it to build a shortlist, then verify filings, debt, and cash flow yourself. Check whether each stock would also pass your basic value filter, and ask why it did not if it fails.

References

BNN Bloomberg, Robot ETF leaves pros in the dust, finding hidden gems: https://www.bnnbloomberg.ca/robot-etf-leaves-pros-in-dust-finding-hidden-gems-1.1155204

Barchart, This ETF Lets AI Do the Stock Picking: https://www.barchart.com/story/news/19386721/this-etf-lets-ai-do-the-stock-picking-has-it-done-better-than-the-sp-500

ETF Central, AIEQ vs SPY: https://www.etfcentral.com/compare-etfs/AIEQ-vs-SPY

Gao, Jiang, Yan, A Test of Lookahead Bias in LLM Forecasts (arXiv): https://arxiv.org/abs/2512.23847

Vanguard, Russell 1000 Growth Index Fund Annual Shareholder Report, August 31, 2025: https://fund-docs.vanguard.com/AR3350.pdf

Vanguard, Russell 1000 Index Fund Annual Shareholder Report, August 31, 2025: https://fund-docs.vanguard.com/AR3348.pdf

FXStreet, Did growth beat value in 2025?: https://www.fxstreet.com/news/did-growth-beat-value-in-2025-202601060516

Was this article helpful to you? Please tell us what you liked or didn't like in the comments below.

About the Author: Chanuka Geekiyanage

Chanuka is a content writer, editor and contributor at Panaprium, specializing in long-form editorial articles. He conducts in-depth research and produces well-structured, carefully edited content with a strong focus on clarity, SEO best practices, and consistent formatting. His work spans lifestyle, fashion, wellness, and technical topics, and all articles are reviewed and refined to meet Panaprium’s editorial standards.

What We're Up Against

Multinational corporations overproducing cheap products in the poorest countries.
Huge factories with sweatshop-like conditions underpaying workers.
Media conglomerates promoting unethical, unsustainable products.
Bad actors encouraging overconsumption through oblivious behavior.
- - - -
Thankfully, we've got our supporters, including you.
Panaprium is funded by readers like you who want to join us in our mission to make the world entirely sustainable.

If you can, please support us on a monthly basis. It takes less than a minute to set up, and you will be making a big impact every single month. Thank you.