Built, not bought

Machine-learning research

Algotoria runs its research as a set of numbered, self-contained experiments, each with its own data, configuration, harness and written record. Work does not graduate because a backtest looked good. It graduates by climbing a fixed evaluation ladder — smoke test, signal triage, single-seed screen, a three-seed reproduction, nested walk-forward at book level, a robustness battery, a shadow window, and finally a pre-registered, one-look forward gate whose pass criteria are written down before the out-of-sample window is ever read.

That last rule is the point of the whole apparatus. The forward window is opened once, against criteria fixed in advance, and the result stands either way. Supporting it are a leak-free data protocol that masks the boundary between training and held-out data, a registry of numbered robustness laws distilled from earlier mistakes, and append-only ledgers of adopted, rejected and pending verdicts — each entry auto-stamped with its date and the exact commit that produced it, so any decision can be reconstructed years later.

The models themselves are built in house. The current line is a multi-scale convolutional network that reads one five-minute bar stream at four horizons at once — five minutes, one hour, twelve hours and seven days — through a purpose-built temporal encoder with cross-instrument wiring. It is now in its fifth generation, and each generation is a documented architectural change rather than a re-tune. The most distinctive piece is not the network but the target it learns: an in-house label that compresses move size and how long the move persists into a single continuous position, computed net of realistic trading costs and under a volatility-scaled stop, so the model is optimising something an account can actually capture.

Alongside price-derived work, Algotoria maintains an active research programme into exogenous signals — order-book liquidity and depth imbalance, social and news-flow sentiment, on-chain activity such as address counts and whale transfers, and capital-flow measures such as exchange net flows and stablecoin supply. These are researched properly and judged by the same ladder as everything else. To date none has earned a place in a production model: the edge has repeatedly proved to live in multi-scale price behaviour itself. We publish that outcome because a research programme that never rejects anything is not a research programme.

Every model is trained on Algotoria’s own hardware, through Algotoria’s own data downloaders, backtester and optimiser. No training is outsourced and no third-party modelling service is used.

Evaluation ladder

Eight stages, R0 to R7, ending in a pre-registered one-look forward gate

Reproduction

Three fixed seeds before any result is believed

Selection metric

Calmar (CAGR ÷ max drawdown), with hard disqualification above 30% max drawdown

Cost realism

Round-trip cost bands calibrated per venue — Binance, OKX and Bybit modelled separately, 14 to 28 bps

Training

In house, on own research hardware, through an own end-to-end pipeline

Algotoria Spike is the product of this research line. It has not launched, no client has ever held it, and its published figures are simulated and carry that marker wherever they appear.

Risk monitoring

The Algotoria Monitoring System is proprietary and deliberately separate from the machinery it watches. It is entirely independent of signal generation and trade execution, connects through its own read-only API keys, and is hosted inside a private cloud network isolated from the public internet. It can observe every account and change none of them — which is precisely what makes it trustworthy as a control.

It tracks more than ten risk parameters for every separately managed account and raises real-time alerts to the NOC and the trading team whenever one crosses a pre-defined threshold. Alerts are graded, and the thresholds sit deliberately below the limits themselves, so a problem is escalated while there is still room to act rather than at the moment the limit is breached.

Monitoring runs against hard portfolio limits rather than opinions. Gross market value is capped at three times account equity, and in practice the daily average sits well below that. Each strategy carries a 95% one-day Value-at-Risk ceiling, no single sub-strategy may contribute more than 5% of that figure, and drawdown budgets are staged — the first stage at two-thirds of the budget triggers an Investment Committee review of the account, the second at the full budget puts the decision to the client.

Underneath it all is comprehensive structured logging across the trading, execution and monitoring estate — every alert, collector run, order and reconciliation recorded in a queryable form, which is what allows any position, fill or alert to be reconstructed after the fact for a client, an auditor or a post-incident review.

What triggers an alert

Position deviation from the master account, leverage above two-thirds of the account limit, drawdown at two-thirds of budget, daily loss beyond 5%, insufficient margin, and API or connectivity errors

Leverage

Hard cap of 3× gross market value to AUM; typical daily average 0.8× to 1.0×

Value at Risk

95% one-day limit of 2.0% on Stable and 2.2% on Diversified; maximum 5% single-contributor share

Isolation

Separate read-only API keys, private network, independent of signal generation and execution

Human cover

A dedicated NOC/SOC on continuous global shifts — 24/7/365

Automated telemetry is the floor, not the ceiling. A dedicated Network and Security Operations Centre runs continuous shifts globally, so an algorithmic anomaly or an exchange API degradation is triaged by a human engineer at any hour. In practice that human intervention is needed roughly once a month.

An AI-native firm

Algotoria is an AI-native asset manager. Beyond the systematic trading engine itself, every operational function — research, software development, operations, client support, compliance and accounting — is executed through AI-assisted workflows under human-in-the-loop oversight, governed by the firm’s Board-approved AI and External Data Provider Policy.

The boundary is absolute and worth stating plainly. Trading is deterministic and non-discretionary. Live signal generation and order execution are governed by algorithms, not by language models. AI sits one layer above the trading loop, in the work that surrounds it.

What makes that layer work is a knowledge base built to be read by machines as well as people. The firm’s wikis carry routing tables that send an agent to the one authoritative file for a question, a standing rule that answers must be grounded in a cited section rather than general knowledge, and an explicit instruction to say “the wiki is silent” rather than guess. On top of that sit detailed playbooks — written procedures for recurring work such as compliance forms, the monthly performance refresh, client-guide verification and wiki maintenance — which an agent follows step by step instead of improvising. The result is that two agents, or an agent and a new employee, answer the same question the same way, from the same source, with the citation attached.

Every consequential decision has a named gatekeeper. Nothing that moves money, changes risk, or reaches a client passes on an agent’s own authority.

Trading strategy

The Investment Committee governs the strategy roster and weights, and gates promotion from incubation to production; a new strategy must additionally clear its pre-registered forward gate plus owner sign-off

Risk limits

Any change to a hard-coded risk limit requires unanimous Investment Committee approval and notification to the Board

Production deployment

Merge-request only, through feature to dev to stage to main; agents work in isolated worktrees and never push to a protected branch; every diff is reviewed by a named human engineer

Compliance and client matters

Account opening, AML case closure and regulator filings require a named human reviewer; the MLRO approves in writing any AI use touching KYC/AML data

Client-facing output

Every client-facing deliverable is reviewed and signed off by a named human before despatch

Governance

Board-approved AI and External Data Provider Policy, aligned to ISO/IEC 42001 and 23894, ISO/IEC 27001 and 27701, the NIST AI Risk Management Framework, the OWASP LLM Top 10 and the EU AI Act

This is also the structural reason Algotoria charges no management fee. AI-native operations lift the productivity of a deliberately lean team far enough that the firm can run on performance fees alone.