Skip to main content

Lab Notebook · Trading Bots Line

TraderBot: Claude reasons about the market instead of following one fixed rule

In May–June 2025 we rebuilt the 2021 crypto-trading-bot idea, but this time an LLM (Claude Sonnet 4) reasons about multiple technical indicators — RSI, MACD, Bollinger Bands — instead of following one hard-coded EMA rule. Sandboxed paper-trading only, never real capital. It proved an LLM can hold and reason about several signals at once, which is a real capability finding — not the same claim as having a durable trading edge.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 15 May 2025
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 15 May 2025
Living document. Reviewed 15 May 2025
Entry date
15 May 2025
Category
Agents
Lead over the world
the proof, not the lead
Access
🔓 Public
ResearchProven

The same trading-bot idea from 2021, rebuilt with an LLM reasoning about several indicators at once instead of one fixed rule — the founder calls the difference night and day.

Claude Sonnet 4
LLM reasoning over market signals
RSI, MACD, Bollinger Bands
simultaneous technical indicators reasoned about
$10,000
sandboxed virtual paper-trading balance

Honest evaluation

Proven

The capability claim — an LLM reasoning coherently over several simultaneous indicators — is proven; a durable trading edge is a separate, explicitly untested claim.

What would prove or disprove it further

What would prove or disprove it further: the durable-edge question this report deliberately leaves open is the real next test — run TraderBot2 against live paper-trading, or a small amount of real capital, over a long enough window and across more than one market regime, and check whether its risk-adjusted returns beat a simple buy-and-hold or the 2021 EMA baseline. Given the 2021 lesson that apparent edges decay fast, that test would need to run considerably longer than one favorable stretch to mean anything.

The evidence — full reasoning behind the verdict

Verdict: proven, on the specific capability claim the report actually makes: an LLM (Claude Sonnet 4) can hold and reason coherently about several simultaneous technical indicators — RSI, MACD, Bollinger Bands together — rather than following one hard-coded rule in isolation. That's a real, working capability, demonstrated in a running sandboxed system, not a claim about trading performance.

The report is careful to draw the line the verdict depends on: "reasons well about multiple indicators" is not the same claim as "has a durable trading edge," and it says so directly rather than letting the two blur together. All trading happened against a $10,000 virtual balance in a sandbox; no real capital was ever risked, and no edge claim is made or implied.

Four years after a fixed-rule trading bot was abandoned for lack of a durable edge, the same question got asked again — this time with an LLM doing the reasoning instead of one hard-coded indicator.

What it was

TraderBot and its successor TraderBot2 — AI-powered crypto trading bots using Claude Sonnet 4 for market analysis, with real-time dashboards, multiple technical indicators, and sandboxed paper-trading against a $10,000 virtual balance.

What we built

Research

A FastAPI backend pulling live data from the Binance API, feeding RSI, MACD, and Bollinger Bands into Claude Sonnet 4 for reasoning about market state, with pandas and plotly driving real-time dashboards. TraderBot2 went further — CLI monitoring and adaptive learning — and is the more mature of the two. All trading happened in a sandbox against a virtual balance; no real capital was ever risked.

What we learned — including the honest negative(s)

The direct comparison to 2021's CryptoTrader is the point of this entry. That earlier bot followed one fixed EMA rule and never found a durable edge. This one has an LLM holding several simultaneous signals in mind — RSI, MACD, Bollinger Bands together, not one rule in isolation — and reasoning about what they mean jointly. The founder describes the difference in capability as night and day.

But say the honest part plainly, because it echoes the 2021 lesson directly: reasoning well about multiple indicators is not the same claim as having a durable trading edge. This was sandboxed paper trading only, against a virtual balance, never real capital — and market alpha is exactly the kind of thing that looks real in backtests and paper trading and decays the moment it meets a live, adversarial market. That open question is not answered by this project, and this report doesn't pretend otherwise.

Where it went / status

Working prototypes — TraderBot2 the more mature of the two, with CLI monitoring and adaptive learning layered on top of TraderBot's core architecture. Neither has traded with real capital. The proven result is narrower and more honest than "a good trading bot": an LLM can hold and reason about multiple simultaneous technical signals coherently, which is a real capability finding worth having on the record on its own terms.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • Sandboxed paper trading only — never risked real capital.
  • 'Reasons well about multiple indicators' is not the same claim as 'has a durable trading edge' — the 2021 lesson that alpha decays fast still applies.
  • TraderBot2 is more mature than TraderBot (CLI monitoring, adaptive learning) but both remain working prototypes, not production trading systems.

Where this connects

Sources

  1. vocabotics project audit — TraderBot / TraderBot2 (Claude Sonnet 4 market analysis, sandboxed paper trading), May 2025vocabotics internal project history · as of May 2025

    We use cookies.