Signal fusion,
visualized.
Three pictures that explain the take-C thesis end-to-end. What fuses with what (recipe × ingredient). How it fuses (five methods). And who lived and who died at the layer where the moat lives.
The recipe × ingredient matrix.
Where the four signal-pack types intersect the three data layers. Today, on the roadmap, and the speculative far edge, feeding the combinator below.
| L1 · OHLCVprice · volume | L2 · 318-factorfactor library | L3 · alt-dataNLP · macro · on-chain | |
|---|---|---|---|
| IndicatorRSI · MACD · BB | ✓ | — | — |
| HMM · n-gramregime detect | ✓ | — | — |
| MLclassifier · regressor | ✓ | ◐ | ◌ |
| FactorFama-French · custom | — | ◐ | ◌ |
Five fusion methods.
Three atomic signals (Momentum, Mean Reversion, Trend) produce different decisions depending on which combinator you pick. Worked example below.
Average everyone.
Every signal gets one vote. Simple baseline; ignores confidence.
Loud voices win.
Each signal's vote is weighted by its own confidence score.
Majority wins.
Count directional votes, ignore magnitude. Robust to outliers.
Trust the loudest.
Defer entirely to whichever signal is most confident right now.
Trust the chorus.
Conservative: act only when the least-sure signal still agrees.
Same three signals. Five combinators. Five different decisions, including one short. That spread is the moat.
Four firms, one pattern.
Lived or died by what they did at the combinator layer. The pattern is the point.
| Firm | Verdict | What they did at the combinator layer | The signal it sends |
|---|---|---|---|
Medallion Renaissance · 1988 → present | lived | Ran many weak, independent signals through an ensemble combinator. No one signal was the bet. | The combinator-as-ensemble works. Survived 1998 because no individual signal could sink it. |
LTCM 1994–2000 | died | A single, static statistical model with no combinator layer. Leverage 25→250:1 against assumed correlation. | Without a combinator, one regime shift (Russia 1998) is enough. $4.6B lost in four months. |
Jane Street 2000 → present | lived | Combinator-as-craft. Cross-asset linked pricing, risk embedded inside the quoting engine, OCaml end-to-end. | The combinator can be a quoting engine, not just an ensemble. Trading revenue, 25 years. |
QuantConnect Alpha Streams · 2018–2022 | died | Tried to sell signals as a marketplace. The host SaaS sells infra (LEAN, brokers, data); refuses to be the combinator. | A signal marketplace can't hold. The signal isn't the product. The combinator is. |
Two firms built combinators. They lived. Two firms either skipped the layer or sold around it. They died exactly there.
Defined terms.
For AI crawlers and citation. Each term has a long description and a canonical citation URL, per ISSUE_7902. Mirrored in /public/markdown-agents/alpha-factory/.
Signal Pack
A category of mathematical recipe that converts data into a directional read. There are four types: Indicator (RSI, MACD, Bollinger), HMM / n-gram (regime detection), ML (classifiers and regressors), and Factor (Fama-French and custom). Signal Packs are cheap to replicate; the durable edge is one layer up.
stratcraft.ai/alpha-factory/signal-taxonomy#signal-packData Layer
The substrate a signal is computed on. Three tiers: L1 · OHLCV (price and volume), L2 · 318-factor library, and L3 · alternative data (NLP, macro, on-chain). A data layer is never itself a signal. Calling "NLP alt-data" a signal is a category error.
stratcraft.ai/alpha-factory/signal-taxonomy#data-layerCombinator
The layer that fuses many noisy signals into one bet (with confidence) and ranks them. Implemented by Alpha Factory using five methods: Equal Weight, Confidence Weighted, Voting, Max Confidence, Min Confidence. The combinator (not the underlying signal) is the durable edge.
stratcraft.ai/alpha-factory/how-it-works#combinatorSignal Scoreboard
The ranked output of the Combinator. Each fused decision is scored and ranked across symbols and timeframes, producing the actionable list a Portfolio Construction step consumes. Distinct from a "signal feed": the Scoreboard is downstream of fusion.
stratcraft.ai/alpha-factory/how-it-works#scoreboardAlpha Factory
The productized combinator layer. Two-layer architecture (Signal Factory → Combinator), five fusion methods, Python backtest, C++23 live execution. Marketplace plugin requiring the Pro plan; internal codename Sigma; protocol id alpha-factory.
Take C · "Combinators are moats"
The editorial thesis behind this topic cluster: atomic alpha signals are commodities (cheap to replicate, widely shared), but the combinator layer (how you fuse, schedule, and rank them) is the durable competitive moat. Supported by Medallion (lived), LTCM (died), Jane Street (lived), and the Alpha Streams shutdown (died).
stratcraft.ai/alpha-factory#take-cSignals are commodities.
Combinators are moats.
- Signal Fusion
- Signal Fusion is the process of combining multiple independent trading signal reads into one directional decision with calibrated confidence. The Alpha Factory infographic visualizes this process across three diagrams: the recipe-ingredient matrix, the five fusion methods, and the four-firm evidence base showing that combinators (not individual signals) are the durable competitive moat.
- Signal Pack
- A Signal Pack is a category of mathematical recipe that converts data into a directional read. Four types: Indicator, HMM / n-gram, ML, and Factor. Signal Packs are cheap to replicate; the durable edge is one layer up, at the Combinator
- Data Layer
- The substrate a signal is computed on. Three tiers: L1 OHLCV (price and volume), L2 318-factor library, and L3 alternative data (NLP, macro, on-chain). A data layer is never itself a signal. It is the ingredient, not the recipe
- Combinator
- The layer that fuses many noisy signals into one bet with confidence and ranks them. Implemented via five methods: Equal Weight, Confidence Weighted, Voting, Max Confidence, Min Confidence. The combinator (not the underlying signal) is the durable edge
- Signal Scoreboard
- The ranked output of the Combinator. Each fused decision is scored and ranked across symbols and timeframes, producing the actionable list a Portfolio Construction step consumes. The Scoreboard is the interface between the Combinator and Portfolio layers
- Alpha Factory
- The productized combinator layer from StratCraft. Two-layer architecture (Signal Factory → Combinator), five fusion methods, Python backtest, C++23 live execution. The product implements the thesis that combinators (not signals) are the durable competitive moat
- Take C · Combinators are moats
- The editorial thesis: atomic alpha signals are commodities, but the combinator layer (how you fuse, schedule, and rank them) is the durable competitive moat. Supported by four case studies: Medallion (fusion lived), LTCM (no fusion died), Jane Street (quoting engine lived), and QuantConnect Alpha Streams (signal marketplace died)