Platform · Market systems & execution
Trading & research platform
A research-to-execution platform for systematic trading. Write a strategy once in a purpose-built language; research it, stress it, deploy it and supervise it on the same compiled logic, so what was tested is what trades.
The problem
Why systematic strategies disappoint once they trade.
Three engines drift
The backtest, the simulator and the live engine are usually separate implementations. Over time they stop agreeing, and nobody measures by how much.
Research rewards fitting
An optimizer always finds a winner. Few tools say plainly when that winner is noise.
Live needs a supervisor
Limits, reconciliation with what the exchange actually holds, and protective orders that follow the strategy's own rules, every bar, without a break.
Claims need checking
Strategies arrive from other tools with impressive results. Their owners need an independent answer: does it reproduce, and is the edge real?
Approach
Six rules the platform is built on.
One strategy, one compiled form
Backtest, paper and live trading evaluate the same compiled logic. There is no second implementation to drift.
Parity is measured, not assumed
One reference defines every fill, stop and cost. The fast engine is tested against it, and running sessions are compared continuously with what the market actually filled.
Deterministic by rule
Nothing that touches money has a default or a guess. A missing price, cost or size stops the run, loudly. Market rules are read from the market, never assumed.
Honest research
Every candidate is judged the same way: out of sample, across many instruments, with costs charged. Too little evidence means no verdict.
One door to the market
Every order passes a single gate that enforces layered risk limits, and every running strategy has a supervisor reconciling expected and actual positions.
An assistant that cannot trade
A built-in research assistant can write, explain and run studies. By construction it has no path to an order, a deployment or a credential.
How it works
From a written strategy to a supervised position.
- WriteIn a purpose-built strategy language, with declared parameter ranges, validated against an instrument before anything runs.
- BacktestMany trials at once on GPUs, charging the market's fees, slippage and carrying costs.
- ValidateOptimization inside the declared ranges, then walk-forward and resampling tests, and the same verdict process for every candidate.
- ResearchA written study of where the result comes from: anatomy, sensitivity and decomposition, each stage traceable to the run behind it.
- DeployBehind deployment gates, to a paper venue running on live market data or to a live account.
- SuperviseEach position reconciled and its protective orders moved every bar, with alerts to the people responsible.
Strategy language
A strategy is a program, and it is checked like one.
Strategies are written in a language of our own, with its own parser, type checker and compiler. A strategy that cannot run correctly is stopped in the editor, not discovered in the market.
Typed and compiled
Every strategy is type-checked and compiled into one vectorized form. The compiler knows how much history each calculation needs before its first valid value.
A fixed catalog
Indicators, mathematics, market-microstructure measures, time functions and order actions, each with a declared signature. Nothing outside the catalog can be called.
Parameters declared in the strategy
Each strategy declares its own parameter ranges and the settings the engine reads, such as stops, leverage, session locks and per-bar stop levels.
Editor and versions
An in-browser editor with completion validates against an instrument and timeframe before anything runs. Every strategy keeps numbered versions with differences between any two.
Research & validation
An optimizer always finds a winner. These tools ask whether it is real.
Batched GPU backtesting
Many trials run at once in custom GPU kernels, split across GPUs or processors by a memory-aware planner, charging the market's fees, slippage and carrying costs.
One reference, every rule
A single reference simulator states what every fill, stop, lock, liquidation, fee and carrying charge means. The fast engine is tested against it.
Optimization with honesty
Searches only the ranges a strategy declares, and reports robustness, sensitivity and whether the best point sits on a plateau or a spike.
Walk-forward and resampling
Walk-forward tests whether parameters hold out of sample across rolling windows. Resampling a run's trades gives the spread of outcomes and the chance of ending ahead.
A consistent verdict
Every candidate is judged the same way: walk-forward across a whole universe of instruments, costs charged, a minimum amount of evidence, and a check for fragile parameters.
Entries measured alone
An exit-free mode measures whether a strategy's entries carry anything before its exit logic decides which luck it keeps.
Where a result comes from
Anatomy breaks a result down by period and contributor. One-at-a-time sensitivity moves each parameter alone. Decomposition tests each part of a strategy separately, against random entries.
Written research
Staged research runs in the background, records every input, and renders as a written report with figures, in an impersonal voice.
Execution & risk
One door to the market, and a supervisor behind every position.
Deployment gates
A strategy version needs a passing composite score before it can deploy, and an account must be verified, with risk policies set and one session per account and instrument.
A paper venue that answers like the market
A simulated account on live market data that behaves like the real venue, including its protective and trailing stops, so a rehearsal exercises the real mechanism.
The order gate
Every order passes one gate: it must be sized, it takes a per-account lock, and it is checked against risk limits. Order management then splits or queues it to fit the venue's limits.
A supervisor per strategy
Each deployed strategy has a supervisor that compares the expected position with what the venue holds, submits corrections, and can be stopped or flattened by an operator.
Live exits that match the backtest
The backtest's lock, session and stop rules are applied live by moving the venue's own protective orders on every closed bar.
Layered limits
Leverage, size, exposure, daily loss, drawdown, order rate and data staleness, each with its own breach action, at strategy, account, portfolio and platform level.
Data, assistant & operations
Everything around the strategy, built to the same rules.
A market-data lake
Historical archives of trades and order books loaded with per-file checks, live feeds in a shared cache, completeness checks, and a refusal to serve any range it cannot serve in full.
Microstructure features
Order-flow and order-book measures aligned to bars, and the market's settled carrying costs charged at every settlement in backtests.
Market intelligence
Movers and volume, regime classification, correlations and carrying-cost comparisons across instruments.
A research assistant with no hands
Tools that call real platform functions and declare their effect. None can place, change or cancel an order, deploy, or touch a credential, and tests hold that line.
Review panels and watchers
Long-running assistant jobs with a full event log, adversarial read-only review panels on a single question, and standing watchers that report and never act.
Shared workspaces
Editor tabs kept on the server and streamed live, so browser tabs and the assistant stay in step and any tab can become a strategy version.
One release door
Every change reaches production through one gate: the full test suite, and a paper-replay parity check whenever execution code changes.
Watched continuously
Every running session is checked on a schedule, and its recent trading is backtested and compared with what the venue actually filled. Alerts go to operators by severity.
Who it is for
Three kinds of user, one standard.
Systematic traders & desks
Write, test and run their own strategies in the web console, or from code through a programmatic client with signed access.
Owners of an existing strategy
Independent due diligence: the strategy is translated, reproduced trade for trade against the owner's own records, researched across its whole history and a universe of instruments, and written up with a funding verdict. The owner then gets dedicated, isolated access to see the work.
Our own research
The house research program runs on the same tooling, under the same rules.
Connections
The same ideas, elsewhere in our work.
Market simulation research
A strategy deserves a faithful market to fail in before it meets a real one. Simulated markets are judged by whether they measure like real days, which is the same demand the platform makes of its backtests: agree with live, or say why not.
Accounting & tax system
Every fill, fee and funding payment ends as a ledger entry. Exact money and records that reconcile to the cent are one discipline, whether the ledger belongs to a strategy or to a company.
Agent runtime & orchestration
Parity between backtest and live can itself be audited by a swarm of scoped agents: one reading each side, one measuring the divergence, two debating it, every step recorded.
Geospatial intelligence platform
Market data and the movement of the world are both streams of events on different clocks. A position, like a price, is only as good as the time it was observed and the feed it came from.
Engage
By invitation.
Every engagement begins with a conversation and is scoped before work begins.
Access
Work with the platform under agreement. By invitation.
Subject: AccessStrategy due diligence
An independent answer on a strategy you hold or are about to fund.
Subject: ResearchPartnerships
Distribution, data or infrastructure partners.
Subject: PartnershipInvestment
Briefings and a demonstration, under confidentiality.
Subject: Fundingcontact@carbonyl.org