Oliver

@ScriptedAlchemy @nstlopez Can I also join the discord ? Currently building my second trading after a first on trend / news follower strategy

@nstlopez Check discord message. I sent you the “master plan” that contains the system end to end. From project creation to deployment

Going full quant and training financial model. Stealing some ideas from TikToks recommendation engine around realtime adaptive training of the model.

@KalleRaus Awesome - currently building an Agent myself- what strategies do you employ?

For 2 years I have been automating what I have learned from 20 years of trading markets. It is now a live portfolio of systematic strategies. Performance so far: Apr 1 2024 -> Apr 28 2026 +98.1%

@0xnfrith @svpino Have you built the orchestrator yourself? I have build one as opencode plugin. Basically eng manager plus engineers can be spawned as teams. It gets very complex quickly with worktree and branch management

@svpino dispatcher/scheduler/orchestration pattern let me run ~3-5 at a time. albit those are spread across codebases but its still multiple agents running at *exactly the same time*

I don’t believe people who say they are running “12 parallel coding agents”. Either they are lying for clicks, or I’m a complete retard who can barely keep up with a single Claude instance.

@Suryanshti777 At least that was my key learning when I implemented an opencode plugin with an agent team

@Suryanshti777 What gets really complex is work management for agent teams. Individual workers need to build on the right branches so that previous work is accounted for correctly and dependencies are resolved. If that's now done well, you will get 5x crappy code

@ScriptedAlchemy Also news cycle is rather hard to backtest... Im not yet at realtime - I first need to get the backtest running for my current strategy running. I was also thinking about EU stocks - there is no SEC EDGAR database, so information is more asymmetrical

My model is trained to cycle options premiums every 2-6 weeks. So it spots horizons. Think this is more reliable than news cycles. Also model connected to socket on alpaca so it has realtime observation Looking and developing a special tokenizer just for stocks which would let me provide huge prompts with lots of data as only a few tokens

I’ve been fine tuning a model locally. Today i broke and got my hands on a h100 instance I can ssh into. Totally worth a dollar an hour.

@ScriptedAlchemy Yeah, I had been trading on 2-5 day horizon. It' had been very dependent on macro news and thus bets. Have been already extending the timeframe now

@ScriptedAlchemy I want to start with 200 mid market, around the 50th percentile of market cap. Not so big that all eyes are on it, but not so small that 8-Ks or news has only limited influence on the stock. For top 100, I would have assumed that most news are immediately reflected in price

@oliver_bauer Right now focus on the top 500 by market cap, 12 years of data. Fine tuned qwen model.

@ScriptedAlchemy So I've reviewed options where I can get a good enough model that is fast with enough parallel requests...

@oliver_bauer Yeah. And news, option chain data, and about 170 features/technicals.

@ScriptedAlchemy Are you using data / text from sec filings / 8-Ks? I am currently working on a post earning announcement drift system,. Parsing all the 8-K, extracting the attachments, and extracting the investor letters is toke intensive 😜

@oliver_bauer Yeah. Evals look promising. Blind backtest as well. Crisis backtest (bad market) all come back profitable.

@nahcrof What is the concurrency limit?

Yeah, and you can get it at the cheapest price on the market using CrofAI (and it’s available on our $5/month plan)

@ParadisLabs @neuralmanifold Haha - so a lot of data wrangling, understood

@oliver_bauer @neuralmanifold Do you count excel as programming? If so, then lots, lol

YTD returns crossed over 500% today. Despite significant macro sell off due to Iran war. And with no margin. Pretty sure I'm outperforming people that charge $5k+ per year for research and technical signals? Some of the big YTD drivers: > AI supply chain bottlenecks make up the bulk e.g. $LITE, $COHR, $AAOI, $AXTI, $TSEM, $IQE, $SOI, $AEHR > $SNDK and Samsung; with rocketing memory prices/supply constraints > Other random momentum trades like $SATL, $SIDU; front running the SpaceX IPO I plan on sharing more of my thought process going forwards like I have this week with some Taiwanese/Space stocks. Still very new to X, but I'm getting more accustomed to sharing my thoughts. Please bear with me! Let me know if you have any questions though, always happy to answer as many as I can. I'm enjoying interacting with you all :) (And no, I do not have any paid discords or signal services, those are all scams. All my ideas and thoughts will be posted right here on @X)

@ParadisLabs That is so awesome! Is that's your full time job? I am current building an LLM-based support analysis tool, but I have not cracked the nut yet

@oliver_bauer Nailed it. Buy and hold higher conviction names, then trade themes/stocks I like the look of but don't have enough conviction in for a longer term holding

@the_proptrader Yes yes yes! This is what I am trying to build, but I am not there yet! I'd not recommend opus - it hallucinates too much for financial markets.

To autonomously backtest a trading strategy, all you require is the following: 1. Historical OHLC Data 2. Karparthy's Autoresearch 3. GLM-5.1/Opus 4.6 4. Trading Strategy Spec Then simply tell it to improve the Sharpe Ratio

@ParadisLabs @neuralmanifold How much programming and Ai are you leveraging? I am a good programmer but not yet a good trader and want to learn

@neuralmanifold Thanks Eli. I'm 100% certain that everyone can outperform the market if they put in the work. Just takes time, effort, and the odd L here and there.

@DeepValueBagger This is sooo cool. I am currently building something similar and extract all the investor letters to compare sentiment quarter over quarter. What strategy do you want to use for trading?

This weekend I'm processing through 2025 Q1, Q2, Q3 10-Q, and 10-K for all of S&P500 companies. It's going to take probably a week to get through it but once I'm done, I'm going to have a complete view how these companies are connected via suppliers/competition/risks, and emerging trends across these businesses.

@ScriptedAlchemy @juristr Ah - I have the opposite approach - I have opencode calling other stuff. Interesting. Any specific signals or strategy you pursue with long/short? I want to try post earnings announcement drift (PEAD) as next one

@oliver_bauer @juristr That was v1 but I do options trading and my MCP would just call opencode wrappers for the tools to summarize some parts etc. current version is using realtime event data and PI harness. I’m now training / fine tuning a model for long/short and call/put

Imagine the amount if money you could make by actually spending it in tokens building useful stuff