Platform
The machinery behind the method.
Every principle on this site is enforced somewhere in code. This is where, and what it runs on.
AI native
Designed, built, reviewed, and operated with AI agents, under the same rules as the code.
Cloud native
Containers, a replayable event bus, and services that heal themselves. Self-hosted today, built to move.
API first
Everything the platform knows is served over an API. The dashboard is just another client.
Data
Built
Features
Built
Measurement
Built
Research
Built
Frozen book
Being built
Forward run
Being built
Capital
Being built
From raw data to capital. The solid stages run today; the outlined ones are designed and being built, and nothing reaches capital until they are.
Architecture
Built so that any part can fail, restart, or be replaced without touching the others.
Publish
Market data
Features
Measurement
Research
Event bus
Durable, replayable, ordered. Nothing calls anything else directly.
Writers
One per table
API
Queries and live streams
Dashboard
Just another client
Replay
New consumers start from the beginning
Subscribe
A durable event bus, not services calling services.
Every component publishes typed events to a durable, replayable stream, and nothing calls anything else directly. A service that goes down resumes from where it stopped, and a new consumer can replay history from the beginning without changing anything upstream.
The dependency graph is the architecture.
Services declare what they consume and what they produce, and the running order is derived from that graph, never maintained by hand. A circular dependency stops the system before any data flows, not after it has corrupted something.
One job per component.
Every service computes, stores, or transports, and never two of those. A component that discovers something never writes it to the database itself; a dedicated writer does. Each table has exactly one writer.
A hot path built for low latency.
Computation runs in memory and never waits on the database. Storage happens behind it, in batches, so a slow disk or a database outage cannot stall a measurement.
Data
Clean, complete data beats a cleverer model on dirty data.
Raw data is permanent.
Every bar from every vendor is kept under its supplier's name, including the ones not used. Corrections are new records, and the series used for research is a rebuildable selection over that raw record.
Missing stays missing.
A gap is never filled with a placeholder or a neighboring value. Warm-up periods are masked, so no measurement starts on data it should not have.
Point in time, always.
The universe, classifications, labels, and settings are read as they stood at the moment being studied, so what is known today never leaks into the past.
In detailHow edge is measured
From one measurement to a claim of alpha, and every place a claim can fail on the way.
- 01
Measure every bar
Feature
A feature is one plain measurement taken on every bar of every security: a level, a change, the change in that change, a z-score, an average. It carries as few assumptions as possible and no view about whether to buy. It is computed whether or not anyone expects it to matter, so nothing is measured only where it was expected to work.
In detail - 02
Take the market out first
Residualization
Before anything is measured, the market and a set of common factors, estimated from the data itself, are removed from the returns being predicted. What remains is the part of a move that belongs to the security.
In detail - 03
Measure prediction
Information coefficient (IC)
The rank correlation between a feature and the residual return that follows it, from one opening price to a later one, so only what could have been traded counts. A useful IC is small, a few hundredths, and that is expected.
In detail - 04
Ask whether it holds
IC stability
IC is measured over separate, non-overlapping stretches of time, because neighboring bars share information. A small IC that is there in every stretch beats a larger one that comes and goes, and confidence intervals come from a block bootstrap that respects how market data repeats itself.
In detail - 05
Know what kind of market it was
Regime
A Gaussian hidden Markov model labels each security's volatility state, such as calm, elevated, or turbulent, using only information available at that moment and never the path that followed. Every IC is measured within a regime, so a measure that works only in calm markets is known for exactly that.
In detail - 06
Correct for searching
False discovery control
Hundreds of features are tested at once, so some will look good by chance. The false discovery rate is controlled (Benjamini-Hochberg), every look at the data is counted, and the best of many is corrected for being the best (Romano-Wolf).
In detail - 07
Is it the market in disguise?
Market-neutral tests
A rising market can make almost anything look predictive. Four separate tests stand between a result and the word alpha, and each one has ended ideas that passed the others.
In detailResidual targets
Prediction is judged on returns with the market and common factors already removed.
Static-tilt benchmark
Every book is compared with a version that does nothing but hold the same average lean. If nothing is left after subtracting it, the result is exposure, not alpha.
Shuffled nulls
The real ranking has to beat many random reshufflings of the same data, or it is chance.
Control groups
When an idea claims a specific cause, the same test runs where that cause cannot apply. If the control behaves the same, the cause is not the source.
- 08
Combine
Combiner and book forecast
Many family members are weighted into one forecast for each security, walk-forward, using only data that came before. Weights follow how reliably each member has predicted, and no single member is the book.
In detail - 09
Confirm once, after cost
Alpha
A frozen book is tested once on data nobody has searched, with trading costs taken out. Only what survives is called alpha, and only alpha is ever given capital. The bar is high on purpose.
In detail
The words, defined
- Edge
- Expected value per unit of risk, measured, and always stated with its sample size and confidence interval. Without those it is a hypothesis, not an edge.
- Alpha
- An edge that is statistically validated, repeatable, and positive after costs on data that was not searched. Not returns, and not outperformance.
- Information coefficient
- The rank correlation between a measurement and the residual return that follows it.
- Feature
- A measurement taken on every bar of every security, in three tiers: atomic, interaction, and theory. Much of the industry calls this an alpha factor; we keep the word factor for what is removed.
- Factor
- A common force that moves many securities at once. The market, peer groups, and statistical components are taken out of returns before prediction is judged; forces such as rates and credit are measured as each instrument's sensitivity.
- Predictor
- A registered rule that turns data into a score for every security, with its direction fixed in advance, its memory declared, and nothing fitted. It is the unit that is tested.
- Regime
- A label for the kind of market a security is in, assigned using only what was known at the time.
- Family
- A few variants of one mechanism, registered before any result is seen.
- Book
- The members, how they are combined, and the rule that turns the forecast into positions. The book, not an idea, is what gets tested.
Reproducibility and governance
Nothing depends on someone remembering.
Every result reproduces exactly.
A result can be rebuilt bit for bit from a snapshot of the data, the code, and the specification that produced it. Every output traces back to its recipe and its run.
In detailNo hard-coded numbers.
Every threshold, weight, and period is a versioned setting with a record of who changed it and why.
In detailOne lifecycle for every idea.
Candidate, shadow, active, deprecated. Only one service can change an idea's status, and nothing can promote itself.
In detailWatched while it runs.
Every service reports its health without being asked, failed services restart in dependency order, and metrics, logs, and traces are collected in one place.
AI native, cloud native, API first
Built the way modern systems are built, and run by the same rules.
AI native.
The platform is designed, built, reviewed, and operated with AI agents working under the same rules as the code, and independent reviews from more than one model come before a design is adopted. The research itself is designed to become agentic, inside the same evidence gates.
Cloud native, self-hosted.
Containerized infrastructure, a replayable event bus, health checks, and scaling driven by measured load. It runs on self-hosted hardware today and is built to move to a cloud without a redesign.
API first.
Everything the platform knows is served over an API: requests for queries and streams for live updates. The dashboard is just another client.