Skip to content

Methodology

WhaleClaw’s methodology is built on one core principle: AI reads the messages, math produces the signal. This separation is what makes the system reliable.

Each of the 128+ KOL agents is a dedicated state machine powered by GPT-4o-mini. When a KOL posts a message, the agent:

  1. Parses the natural language — slang, abbreviations, emojis, context from prior messages
  2. Extracts three fields: directional bias, conviction strength, stance timestamp
  3. Maintains state — if a KOL said “long BTC” yesterday and hasn’t changed their view, the agent keeps them as “bullish” until they explicitly shift

This layer needs AI because human traders don’t write in structured data. They write “heavy long here, LFG” or “I’m out, this looks toppy.” The AI understands what they mean.

Layer 2: Deterministic Aggregation (produces the number)

Section titled “Layer 2: Deterministic Aggregation (produces the number)”

Once every KOL’s stance is extracted, the aggregation is pure math:

Composite Score = Σ (bias × conviction × tier_weight × time_decay)

No LLM touches this step. The weighted votes, time decay curves, and conviction multipliers are deterministic calculations. This means:

  • The final consensus number cannot hallucinate. It is exactly what the weighted data says.
  • The calculation is reproducible. Same inputs always produce the same output.
  • Errors are traceable. If the signal seems wrong, you can trace it back to which KOLs drove the score.

Each KOL agent operates as a finite state machine with three states:

StateDescriptionTransitions
BullishKOL is directionally longExplicit bearish/neutral statement
BearishKOL is directionally shortExplicit bullish/neutral statement
NeutralKOL is flat or undecidedExplicit directional statement

Key design decisions:

State persists until explicitly changed. Silence doesn’t mean neutral. If a trader said “long BTC” and then went quiet for 12 hours, they’re still long. Only an explicit reversal changes their state.

Time decay handles staleness. A 12-hour-old call still counts, but less than a 2-hour-old call. This is handled by the weighting math, not by changing the state.

Conviction is extracted, not inferred. “I’m all in” maps to strong conviction. “Maybe long here” maps to weak. The AI makes this classification; the math applies the multiplier.

The Composite Read doesn’t rely on KOL consensus alone. It combines four independent data streams:

SourceWhat it capturesWeight in composite
KOL ConsensusWhat 128+ traders thinkPrimary directional signal
OrderflowWhat the tape showsConfirmation / divergence
Market StructureWhat the chart saysLevel-based context
Macro ContextWhat’s happening in the worldRisk/catalyst overlay

When all four agree, confidence is high. When they diverge, the system either reduces confidence or flags specific conflicts.

In trading, false precision kills. If your signal source can produce different outputs from the same inputs (as any LLM can), you can’t trust it consistently.

WhaleClaw’s architecture guarantees:

  1. Consistency — The same market state produces the same signal
  2. Transparency — Every output can be decomposed into its inputs
  3. No hallucination — The aggregation layer is pure math
  4. Traceable errors — When the signal is wrong, you can identify why

The AI layer does what AI is good at (understanding messy human language). The math layer does what math is good at (producing reliable, reproducible numbers).

See the proof

Walk-forward backtest results across 588 signals from 6 profitable KOLs.

Backtest Results →