Islam Baraka

sharing knowledge related to Fintech engineering over

[Islam Baraka @IslamBarak90]

@theofinr At this point, Me!

[Islam Baraka @IslamBarak90]

@Rasmic everyone!
I even have a video on youtube doing a historical benchmarking over the model

[Islam Baraka @IslamBarak90]

@marcelpociot @typesafeai on a historical backtest benchmarking i ran i did just fine with around ~3.5% profits and > 85% no trade decisions, however there is some contradictory decisions you can see on my video on youtube here
Jev AI Trading Model Tested: 300 Decisions, Real Benchmark Results

[Islam Baraka @IslamBarak90]

I ran TypeSafe’s new Jev model on 300 real trading decisions.

NVDA. JPM. XOM. Bitcoin. Gold.

85% of the time it said: stay flat.

The 46 trades it did take: 23 wins, 23 losses. +3.82%.

Then the report got ugly.

Every single short came back “not applicable” for entry, take-profit, and holding period.

Trades with no internal contradictions: +6.36%
Trades that contradicted themselves: −2.54%

Highest-confidence trades were worse.

I replayed every answer in a controlled backtest lab. Full method, every cutoff, every contradiction — in the video.

Which model should I put through the same 300-decision gauntlet next?

https://t.co/gitO1pUXWG

[Islam Baraka @IslamBarak90]

I just finished a 300-decision trading benchmark with Jev — and the result is much more interesting than the +3.82% return.

5 markets.
300 point-in-time decisions.
46 actual trades.
85% of the time: NO TRADE.
Results:
• +3.82% net return
• 50% win rate
• 1.06 profit factor
• 30 longs / 16 shorts
• 23 winners / 23 losers
At first glance, +3.82% doesn’t look spectacular.
But that’s not what caught my attention.

Jev is fundamentally different from the LLMs I’ve tested in this trading arena.
It doesn’t generate a long analysis and then convert that text into JSON.
It directly returns typed, probabilistic decisions.
That makes it extremely interesting for something like real-time algorithmic trading:
market state → structured decision → execution
without needing a large natural-language reasoning layer in between.

And there’s another fascinating result:

Jev stayed flat on 254 / 300 decisions.

It wasn’t constantly trying to invent a trade.

That kind of selectivity is something I’ve struggled to get consistently from general-purpose frontier models.

The benchmark also uncovered contradictions inside Jev’s answers — including cases where it traded despite:

• low setup probability
• recommendations to stay flat
• missing stop-loss / take-profit answers

Which is exactly why I’m interested in this experiment.

The model isn’t “solving trading.”

But it may represent a very different architecture for building machine-native decision systems.

I’m going much deeper into the results next.

Particularly:

confidence calibration

decision consistency

latency
abstention / NO_TRADE behavior
contradiction rates

comparison against frontier LLMs

This one surprised me.

Jev may be more interesting as a decision engine than as a chatbot.
Going to upload a performance video really soon!

thank @CompleteSkeptic for this great model

[Islam Baraka @IslamBarak90]

@typesafeai How to access it ?

[Islam Baraka @IslamBarak90]

@shivangitech_97 Many reasons, More expensive less token session limits on 100 plan and limited fable useage so you need to go for opus more often!

[Islam Baraka @IslamBarak90]

@zavxai China 💯

[Islam Baraka @IslamBarak90]

@DanielSmidstrup Astra

[Islam Baraka @IslamBarak90]

@pcshipp make sure you closed your chatgpt (codex) not only the close it but make sure its totaly closed even from the system try and then re-open it

[Islam Baraka @IslamBarak90]

@theo On a 100$ plan one prompt perfect 12% of weekly limits!
Yet to be discovered on another task that usually consume 4-6% of Soal at max, going to update this with the final result and increase

[Islam Baraka @IslamBarak90]

@pradeepXkapoor Totally agree!

[Islam Baraka @IslamBarak90]

@tianzige4 Everyone from China is most welcomed in Egypt!

[Islam Baraka @IslamBarak90]

@LiuXininBeijing Welcome! Glad to see you visiting Egypt

[Islam Baraka @IslamBarak90]

@milesdobrenski @OpenAI Congrats!

[Islam Baraka @IslamBarak90]

@buhaimedi Too many projects are kicked off this way now days, Well said

[Islam Baraka @IslamBarak90]

@shen_shiwei Since you're in Cairo now, let me know if you have time to grab a coffee.

[Islam Baraka @IslamBarak90]

what the hell! Claude Fable 5 fallback to opus 4.8 on a ui ux review task
what is wrong with you @ClaudeDevs @claudeai

[Islam Baraka @IslamBarak90]

@Henryf1w Not so true, as its consume more tokens for lower quality that's the main reason, Its a very good Jump however not the Fable 5 comparable yet

[Islam Baraka @IslamBarak90]

@LeoBuilds_ Grok 4.7, That said its a very good model BTW.