Claude Fable 5 for Trading (it's getting there)

Claude Fable 5 built a Tom Trades Candle Behavior Reversal (CBR) indicator in 35 minutes — first model to do it in one shot.

## What Worked For 3 months, every new model Ron tested failed to mathematically express an "overextension" — price moving one way for 20+ minutes without a pullback. Claude Fable 5 nailed it on the first prompt, generating a v1 indicator with the correct logic, then a v2 that draws the overextension box. Ron frames it as the closest thing to AGI he's seen for indicator work: "It's legitimately thinking instead of the usual guessing the token sequences."

## What Still Breaks The model flagged a 21-minute overextension but missed a valid entry. The cause: the advanced CBR rule is that the breaking candle only needs to break the *previous candle's* high, not form a clean "lower high." Fable 5 defaulted to a mechanical "lower high" definition instead. It's the gap between human discretion and deterministic math — fixable, but it cost a trade signal.

## Token-Cost Lesson Feeding all ~50 of Tom Trades' YouTube videos caused the model to hallucinate and fall back to transcript-fetching instead of frame extraction. Ron now feeds it exactly one video, demands frame by frame extraction, and had Fable 5 process a 9-minute video in roughly 5 seconds of internal work (~100 frames). The third takeaway: keep v1, build v2 separately so you can audit the logic diff.

## The 4-Phase Plan 1. Price-action indicator (this video — done at v1/v2). 2. Backtesting engine — preferably LuxAlgo, or Fable 5 builds one. 3. Order-flow integration using taker volume, likely on MMT (basic plan is $200/month, pro is $4.99/month — "the API part is very expensive"). 4. Live execution with real capital — Ron will start with $1,000, not the $100K flex accounts he sees hyped on X.

## Caveat on Hype "If they're going to make a million dollars in a short amount of time, they either have to win very frequently, which is not really possible anymore because of the fees, or you have to bet a lot, and that is very dangerous." TradingView also has no native MCP/API, and his account got banned during an Opus 4.8 test, which is why he runs this on Cursor instead of Claude Code.