benali ghita

✌🏻

[Chainswatch @chainswatch]

https://t.co/KikG7U4t23

[Chainswatch @chainswatch]

Your risk model might be quietly costing you more money than you think.

Not because you made bad calls.
Because the scenarios your risk model runs to protect your portfolio are partly fiction.

There is only 1 dataset of the S&P 500. 6,064 trading days. That's it.

Fake correlations
When generative AI produces thousands of future scenarios from that single history, it fills the gaps.

With invented correlations and volatility spikes that look plausible but are mathematically wrong.
The model has no idea what it doesn't know, so it makes it up...confidently.

That's where the money goes.

Maximum Entropy
Maximum Entropy was written down in 1957 by a physicist named Jaynes. Around one idea: don't invent what you don't know.

Commit only to what the data tells you. Stay uncertain about everything else.
The catch: it was so slow and unusable at real scale. Days of compute per run. So the industry moved on and accepted hallucination as the cost of speed.

Maximum Guided Diffusion
In February 2026, researchers at ENS Paris, NYU Courant, and Capital Fund Management cracked it.

They built Moment Guided Diffusion “MGD”.

MGD runs at modern AI speed. But at every single step of generation, it enforces one hard rule: match the real data's statistical fingerprint. Nothing invented beyond that.

No fake correlations. No invented structure.
They tested MGD on 24 years of S&P 500 returns, turbulence simulations, and dark matter maps. The fat tails, the volatility bursts, the long-range dependencies all reproduced. Nothing added.
MGD is what honest scenario generation looks like.

This matters for traders who think their risk tools are “good enough” in violent markets.

[Chainswatch @chainswatch]

https://t.co/XHHJOVKMWM

[benali ghita @gbd___]

@chainswatch Not me finding out that volume doesn’t mean healthy market

[benali ghita @gbd___]

@chainswatch The real question is why this still isn’t a priority for more teams