Why Math-Only AI Stock Predictors Will Always Fail

The Experiment That Proves the Point

A PhD-holding software engineer recently documented his year-long effort to build an AI stock predictor. The setup: a Long Short-Term Memory (LSTM) neural network — the same architecture behind early speech recognition — fed a growing stack of input parameters to predict tomorrow’s stock value for over a thousand companies.

His approach is representative of how most people think AI should be applied to markets. Let’s walk through what he tried, what happened, and why it keeps not working.

Round 1: Fundamentals + News

He started with the obvious stuff. Company fundamentals, earnings reports, news sentiment scores, insider buying data. The model ingested all of it and tried to predict short-term returns.

It didn’t work well. Fair enough — those are the inputs everyone uses. No edge there.

Round 2: Add the VIX

A commenter suggested adding human emotion. So he added the VIX — the market volatility index — as a proxy for fear. Calculated 10, 20, and 30-day rolling averages. Retrained the model.

Result: the model basically stopped buying anything. It got almost no buy signals at a 50% confidence threshold. The feature importance analysis showed it put “a ridiculous amount” of weight on intraday range and almost nothing on everything else.

His own words: “The model isn’t really learning.”

Round 3: Add Fear and Greed

The VIX describes volatility, not emotion. So he added the CNN Fear and Greed Index — a composite score that more directly measures market sentiment. One problem: public data only goes back to 2011, which “massively cuts off” the training set.

Results were slightly better. But then he discovered a deeper problem: some stocks are correlated with fear (Tesla drops when people get scared) and some are anti-correlated (Dollarama rises when people get scared). Feeding fear data without a correlation factor gives the model contradictory training signals. It can’t learn because the same input produces opposite results depending on the stock.

Round 4: Add a Correlation Sensitivity Factor

So he added another parameter — a 60-day backward-looking correlation between each stock and the fear/greed index. Positive for stocks that move with fear, negative for stocks that move against it.

Finally, he got slightly better back-test returns than the S&P baseline.

And then he said this, on camera:

He’s Asking the Wrong Question

Here’s where it gets interesting. The problem isn’t his code. The problem isn’t even his data. The problem is his question.

“What will this stock’s price be tomorrow?”

That’s the question his entire system is built to answer. And it’s the wrong question.

Price moves on tweets. On a CNBC segment. On a single fund manager having a bad day. On a Reddit thread that goes viral at 2am. On a hedge fund unwinding a position because a different investment went sideways. Price is noise — the aggregation of millions of short-term emotional decisions by people who are, by their own admission, driven by fear and greed.

You can’t predict noise. You can only try to react to it faster than everyone else. And there is always — always — someone faster. High-frequency trading firms spend hundreds of millions on microwave towers and submarine cables to shave microseconds off their reaction time. You’re not going to out-speed them with a Python script running on your laptop after your day job.

The right question is completely different:

Price and value are different things. This is not a new insight — Ben Graham said it in 1949. But in the age of AI disruption, the gap between price and value is about to get very, very wide across dozens of industries simultaneously. And math-only AI predictors are structurally incapable of seeing it coming.

What Math-Only AI Uses vs. What Actually Matters

Let’s get specific. Here’s what his system — and most AI stock predictors — look at, compared to what actually drives industry-level disruption:

Every single one of these inputs is backward-looking. They describe what has already happened and assume the pattern continues. During normal market conditions, that assumption is... okay. Imperfect, but workable.

During a period where AI is simultaneously restructuring 28 industries? That assumption is catastrophic.

The Reasoning Problem: Why Math Alone Can’t Do This

OK. This is the part that makes people’s eyes go wide.

His LSTM neural network takes numbers in and produces a number out. A predicted return. That’s it. Ask the model why it bought a stock and it can’t tell you. It found a statistical pattern in the training data and assumed that pattern will continue. No explanation. No reasoning. No causal chain.

Now think about what’s actually happening in markets right now.

AI disrupted drug discovery. Pharma can now screen compounds 100x faster. That’s a pharma story, right? A math-only model sees pharma inputs changing and adjusts its pharma predictions.

But here’s what the math can’t see:

• Faster drug discovery changes insurance pricing.More drugs, faster approvals, different risk profiles. Actuarial models get rewritten.
• Changed insurance pricing restructures healthcare delivery.New drugs shift the cost curve. Different procedures become viable. Hospital systems reorganize.
• Restructured healthcare changes hospital construction.Fewer surgical suites, more outpatient facilities. Real estate investment trusts reprice.
• Changed construction shifts materials demand.Different building types need different materials. Steel, concrete, medical-grade plastics all shift.
• Materials demand reprices energy.Manufacturing changes alter industrial energy consumption patterns.

One AI disruption in drug discovery. Five industries affected. And that’s one chain out of 170 we’ve mapped.

No neural network can discover these chains from historical price data. They don’t exist in the numbers. They exist in the reasoning about why one industry’s transformation forces changes in another.

The Treadmill: Adding Complexity to the Wrong Model

Watch the pattern in his development process:

1. Fundamentals didn’t work. Add news sentiment.
2. News sentiment didn’t work. Add VIX.
3. VIX didn’t work. Add fear and greed.
4. Fear and greed contradicted itself. Add correlation sensitivity.
5. Slightly better. “Doesn’t prove it works.”

He’s never asking: “Am I asking the right question?”

He’s building increasingly complex machinery to read tea leaves faster. Each failed round gets patched with another input. The model gets more parameters, more training time, more complexity — and marginally better results on data it’s already seen.

This is the classic optimization trap. You can spend forever getting 2% better at the wrong thing. Or you can step back and ask whether the thing itself is what matters.

This applies to every sector. Every correlation. Every sensitivity factor. The relationships between stocks, between industries, between fear and price — all of them were learned in a pre-AI world. And they’re being recalculated by reality right now, whether the model updates or not.

What Forward-Looking Analysis Actually Looks Like

Here’s the opposite approach. Instead of predicting tomorrow’s price with yesterday’s data, map the structural forces that are reshaping industry value over 1–20 years.

That requires 8 distinct analytical dimensions — most of which can’t be reduced to a number:

Each of these dimensions produces a score. But the score comes from reasoning, not from statistical regression. It comes from asking: “Given what we know about human psychology, geopolitical constraints, historical precedent, and cross-industry effects — what happens to the value of this industry over the next 5 years?”

That’s a fundamentally different kind of analysis than “given the last 60 days of fear/greed correlation, what will the price be tomorrow?”

The Honest Part

Let’s be clear about something. Nobody can predict stock prices. Not us, not him, not Renaissance Technologies, not anyone. Anyone who says otherwise is selling something.

But there’s a difference between predicting prices and understanding structural shifts.

You don’t need to know whether Pfizer will be up or down next Tuesday. You need to know whether the entire pharma industry is being restructured in ways that change its relative value versus insurance, versus healthcare delivery, versus biotech, versus medical devices — over years, not days.

That’s knowable. Not with certainty, but with informed reasoning that’s directionally correct far more often than a math model guessing at tomorrow’s noise.

Related reports: How We Score 28 Industries  |  What Wall Street's $40 Trillion Runs On  |  The Human Bottleneck