Real Production Infrastructure Case

Algorithmic Fleet
Eight trading bots under one shared discipline

noches is not a strategy. It is a fleet. Eight bots run deliberately different logic across four exchanges, and none of them is the authority on its own. What separates this from "a trading bot" is not how fast it decides — it is how reluctant it is to accept a decision. Nothing gets capital until it has survived data it was never built on.

Out-of-Sample Validation Fleet-Level Risk Budgeting Daily Oversight Layer
Design a Trading Architecture More Case Studies

Why This Fleet Operates at a Higher Level

No single point makes the decision. Each bot runs its own logic, but all of them pass through one shared gate — and that gate, not the cleverness of any one bot, is what keeps the system standing.

Deliberate Diversity

The bots are not alike, and should not be: one runs reinforcement learning, one is cross-sectional market-neutral, several trade statistical patterns. When the logic differs, the mistakes do not arrive together.

The Out-of-Sample Gate

Every idea has to hold up on data it has never seen. Overfitting is caught and rejected by statistical testing, not by admiring a good-looking backtest curve.

Failure as a Valid Result

When a model layer measured essentially zero on unseen data, it was retired. Not defended, not re-tuned — replaced by simpler logic that could be justified. It was the hardest call in the project.

The Oversight Layer

An agent reads the fleet's record every day and issues risk guidance: where to size down, where not to trade at all. Its call is recorded and later scored against what actually happened.

End-to-End Architecture

The system is layered so any one layer can be rejected without the rest collapsing. Each night's output is the next night's input.

1

Data Ingestion

Multi-exchange, multi-timeframe streams kept synchronized, with continuous data-quality control.

2

Edge Discovery

A nightly search for durable patterns across the whole tradable universe, not just the current portfolio.

3

Statistical Gate

Every candidate must clear out-of-sample testing. What fails never reaches a bot.

4

Allocation

Capital and portfolio rotation follow what actually survived, with hysteresis so the book is not churned nightly.

5

Risk & Execution

Risk budgets, leverage ceilings, recomputed exits and drawdown guards at fleet level.

6

Oversight Loop

Execution-quality analysis, daily review and retraining; every cycle sets the parameters of the next.

In practice, not on a slide

Out-of-Sample Disciplineacceptance is judged on unseen data, never on backtest beauty
Regime & Event Awarenessrisk sizing adapts to market regime and high-risk events
Execution Qualitythe gap between the best available price and the real exit is measured and corrected
Fleet-Level Riskcorrelation between bots carries a budget too, not just per-bot risk

Continuous Optimization and Operations

This is where most of the work happens. Every night, with no manual step: fresh data arrives, patterns are re-tested, the portfolio is updated, and whatever cleared the gate lands on the bots.

The Nightly Review

  • Fresh data and pattern re-testing across the whole tradable universe
  • Validated output applied automatically, with no service interruption
  • Periodic deep optimization, only for the layers that genuinely benefit from it

Portfolio Rotation

  • Pairs enter and leave on measured fitness, not on conviction
  • Hysteresis, so a small wobble does not churn the whole book
  • Forced eviction of what has stopped working, however good its record was

Operational Risk Controls

  • Daily trade caps and enforced spacing between entries
  • Drawdown guards at both single-bot and fleet level
  • Automatic halt when several bots come under pressure at once
soodo

The fleet's signal output also surfaces on soodo, a separate product that makes these same decisions followable for an end user. The same discipline, seen from the side that only gets to watch the result.

My Role

  • Designed the fleet architecture, from pattern discovery through execution and risk guards
  • Built the statistical gate that decides what is allowed to go live at all
  • Designed the nightly loop: review, portfolio rotation and automated deployment
  • Turned research code into operational infrastructure with monitoring, execution analytics and an auditable record
  • Retired the layers that did not work — the hardest and most valuable part of the job
Book a Trading Architecture Session Back to Case Studies