ludoonchart

this is the most DANGEROUS thing on the whole internet right now the tools Goldman Sachs, Point72 & Man Group built for their own desks, plugged into Opus 5.5 = a 24/7 QUANT DESK that runs on your laptop i plugged them into Claude Code & built a COMPLETE AI trading setup from scratch my setup pulls real market data, writes the strategy, sweeps 212 combos, picks the best one & marks every LONG and SHORT on the chart on its own (ZERO black box, the code & data stay mine) It's built on gs-quant (13.9k stars, Goldman Sachs), csp (Point72's stream engine) and ArcticDB (Man Group's DataFrame database) here is the EXACT build from scratch: 1. Install Python 3.10+ and Claude Code. That's the only prerequisite. 2. Open Claude Code and paste: "Create a fresh venv and install csp from github.com/Point72/csp, gs-quant from github.com/goldmansachs/gs-q… and ArcticDB from github.com/man-group/ArcticD…." 3. Paste: "Get the last 30 days of 1-minute bars for any asset from a public exchange dataset (i used BTC/USD), resample to 15m and store them in a local ArcticDB library." 4. Paste: "Read the bars back from ArcticDB. Compute EMA, Bollinger Bands, RSI and a volume profile with gs-quant's timeseries functions." 5. Paste: "Use csp to replay the bars as a live stream. Write an EMA cross strategy that goes long and short, with 2 bps fees." 6. Paste: "Sweep fast 4-24 and slow 12-96, 212 combos. Rank by Sharpe. Split the 30 days into 4 windows and compute each window's equity curve. Tell me which combos are a plateau and which are a spike." 7. Paste: "Draw every long and short on the chart at the exact bar, and make the agent log fire on the same bar." 8. Remember: the best combo was picked on the same data it's shown on. It's a screen, not proof. Run walk-forward before you trust it 9. Paper trade it for weeks before you even think about live. Check ArcticDB's license before using it commercially links to all three repos are at the bottom be one of the first to use this before everyone else does:

AI labs are f*cked again... Jane Street just co-authored a paper on the problem everyone's scared of: we're running out of training data the idea: take real web text, have a small model rewrite it as math explanations, FAQs, tables and tutorials, then train on the mix here's what they found: ~ the more compute you have, the more rewrites you want: under 10% at small budgets, 32% at the big one ~ their formula predicted the 22B model's loss within 1.1% ~ on math, training on rewrites alone cut the loss from 1.33 to 0.80 ~ the model beat the model that rewrote its data on 10 of 16 domains ~ real data alone started overfitting after 3 passes with Percy Liang from Stanford + Princeton find the paper link below ↓ the data wall might just be a rewrite away

I can't believe these 2 GitHub repos are still FREE - they should be f*cking BANNED… One reads the market for you every day and spits out a decision dashboard. The other gives your AI agent a brain that never forgets Put them together and every daily report becomes a linked note in your vault, so your agent remembers every call it ever made. No bullshit Repo 1: obsidian-mind (5.5k stars) A self-organizing Obsidian vault that gives AI coding agents persistent memory. Works with Claude Code, Codex CLI and Gemini CLI Repo 2: daily_stock_analysis (67.6k stars) LLM-powered multi-market stock analysis: multi-source market data, real-time news, a decision dashboard and automated notifications. Runs on a schedule, cost-free here is how I wired it: step 1 → clone both repos step 2 → open Claude Code inside the vault and tell it to wire daily_stock_analysis into obsidian-mind and save every daily report as a linked note step 3 → pull the daily closes for TSLA, SPX and SPCX and score each one against its 5-day average: BUY above +0.4%, SELL below -0.4%, HOLD in between step 4 → every report becomes a note in brain/daily, and each ticker note gets a new line with a [[wikilink]] back to it step 5 → the hooks do the rest: after every write they validate the note and update the backlinks, before compact they back up the session log, at session end they save a summary to brain/ step 6 → the session-start hook loads your goals and lessons, so every new run starts with memory instead of from zero step 7 → put it all in one terminal: the daily analysis on top, the vault graph and memory hooks below, Claude Code at the bottom links to both GitHub repos are in the comments below take it before someone else does ↓

I'm begging you to use these 2 GitHub repos - they're a f*cking CHEAT CODE One makes AI trading agents rewrite their own prompts. The other throws 1,600 simulated traders at the same chart Put them together and you get a loop that only keeps a change if Sharpe actually goes up. No GPU. No bullshit Repo 1: atlas-gic (2.4k stars) Self-improving trading agents, Karpathy-style autoresearch. The agent prompts are the weights, Sharpe ratio is the loss function. Their README: 54 prompt changes tried, 16 kept Repo 2: MiroFish (84k stars) Swarm intelligence engine. Spawns a crowd of agents and watches where they go here is how I wired it: step 1 → clone github.com/chrisworsey55/atl… and github.com/666ghj/MiroFish step 2 → open Claude Code inside atlas-gic and tell it to wire the MiroFish swarm into the ATLAS rewrite loop, keeping a prompt change only if Sharpe improves step 3 → let the swarm simulate 1,600 paths 60 steps ahead on the latest TSLA 15m bars and read out bull %, p10, p50 and p90 step 4 → feed that swarm consensus into the ATLAS macro layer as a signal for the agents step 5 → every cycle ATLAS finds its worst agent and rewrites its prompt. Sharpe up = git commit, Sharpe down = git reset --hard step 6 → keep it paper trading only so nothing touches real money step 7 → put it all in one terminal: the swarm on top, the agents and the Sharpe landscape below, Claude Code at the bottom be one of the first to use this before everyone else does ↓

I am begging you to use these 2 GitHub repos - they're f*cking INSANE... one repo gives you a full open-source Bloomberg-style terminal, the other turns Claude into a team of 4 value investors that ARGUE with each other i replayed 546 REAL 15-minute TSLA candles & let the agents trade it on paper from scratch my setup reads every bar, debates bull vs bear, and only pulls the trigger when the team agrees (11 TRADES, +$1,996 PAPER P&L) here is the EXACT build from scratch: 1. open Claude Code in an empty folder & paste: "clone github.com/Fincept-Corporati… and explain the equity research screen in five lines" 2. paste: "install github.com/xbtlin/ai-berkshi… and run /investment-team on TSLA. show me what each of the 4 lenses says" 3. paste: "load 21 sessions of TSLA 15-minute OHLCV into a CSV, real bars only, no made-up candles" 4. paste: "replay the bars at 2.85 candles per second with a 58-bar sliding window, each candle grows from open to close like a live feed" 5. paste: "add a 9-bar moving average. when price crosses it after 4 bars on the other side, raise a SIGNAL" 6. paste: "before any trade, the 4 agents vote: business quality, moat & valuation, inversion & risk, long-term. the team lead makes the final call" 7. paste: "every signal the team approves becomes a PAPER trade, 50 shares. mark BUY and SELL right on the chart" 8. paste: "log every thought with the bar's timestamp: why it bought, why it sold, what it learned when it got stopped out" 9. paste: "put it all in one terminal: chart on the left, the agents thinking in a Claude Code window on the right, live P&L at the bottom" be one of the first to use this before everyone else does ↓

JPMorgan just built a ChatGPT for the stock market and NOBODY is talking about it not for chatting. for TRADING the biggest bank on earth trained an AI on 19 BILLION tokens of real trade data. every buy, every sell, every order, across 9,000+ US stocks and now it can generate a fake stock market that behaves like the real one here's how they did it: step 1: collect tick-level trade data from 368 trading days. every single event across 9,000+ US equities step 2: turn every trade into a "word". price distance, size, timing, buy or sell, all packed into one token. same trick ChatGPT uses for text step 3: train a 524M parameter transformer to predict the next trade. on just 3 GPUs step 4: plug it into an order book simulator. the AI places orders, the book reacts, the AI sees the result and keeps going and the results are f*cking insane: > it matches real price behavior 2-3x better than the classic quant models (Hawkes processes) banks have used for years > it recreated the stuff that breaks most models on its own: fat tails, volatility clustering, the way crashes bunch together > simulated fills matched real fills with 0.91 correlation > trained ONLY on US stocks, then thrown at China and Japan with zero retraining. it still held up > you can inject 10x the orders into a stock and watch how the market reacts. a stress test that never happened wall street doesn't need more data anymore it can just generate the market save this, full paper below ↓