I Gave 9 AI Models the Same Stock IPO Data. They All Got It Wrong

#python #ai #machinelearning #discuss

I'm an ordinary guy in China. I got lucky in an IPO lottery —
1,000 shares of CXMT (688825), China's first domestically-produced
DRAM memory chip maker. Cost: ¥8,660.

Before listing day, I ran an experiment.

The Setup

I gave 9 different AI models — Claude, GPT-4o, Gemini, and others —
the exact same data packet:

CXMT's financials (still loss-making as of H1 2026)
Sector comparisons (vs Samsung, Micron, YMTC)
IPO mechanics (A-share STAR Market, no price limit first 5 days)
Market sentiment indicators

One standardized prompt. Nine models. One question: what price
range do you predict for the opening?

What Actually Happened

Metric	Value
IPO Price	¥8.66
Opening Price	¥49.50
Intraday High	¥55.03
Close	¥49.00
Day 1 Gain	+465.82%
Market Cap	¥3.28 trillion (~$455B USD)

For reference: Intel's current market cap is approximately $90B USD.

A company that won't turn profitable until at least H2 2026
just listed at a valuation larger than one of the founding
fathers of the semiconductor industry.

All 9 AIs underestimated this. Significantly.

Why the AIs Were Wrong

This is the part I find genuinely interesting as someone who
uses AI tools daily.

The models reasoned correctly about the fundamentals:

Loss-making company ✓
High valuation vs peers ✓
Uncertain demand cycle ✓

What they couldn't model:

1. National narrative as a price driver

CXMT isn't just a stock. It's China's answer to the US chip
embargo. Retail investors weren't buying earnings — they were
buying a symbol. No training data fully captures that.

2. Zero-float dynamics

On STAR Market IPOs, the tradeable float on day one is tiny.
When retail demand is enormous and supply is near-zero, price
discovery goes vertical. The math is simple but the magnitude
is hard to predict.

3. Crowd psychology at full velocity

AI models are trained on markets that behave rationally most
of the time. They have no good model for the specific fever
of a landmark national IPO.

I sold my full allocation on listing day. The S markers on
the chart are my actual exit points. I didn't hit ¥55 —
nobody does. But I exited profitably.

The money goes toward my son's school tuition. That was
always the plan.

The Actual Lesson

I trusted AI completely. What I underestimated was
human frenzy.

This isn't an argument against using AI for investment research.
I use AI tools extensively for stock analysis — they're excellent
at processing financials, comparing peers, identifying risks.

But there's a category of market event where the key variable
is how many people believe something matters — and that's
precisely what language models struggle with.

Use AI as one input. Not as an oracle.

I document my real-money AI experiments (including failures)
at ordinarymantrying.com.
Ordinary Chinese guy, real stakes, no filter.