A quote arrives, ages, and is gone. By the time a model has finished reasoning about it, the book has moved, the spread has widened, and the order it recommended is a bet on a market that no longer exists. Most trading systems are not wrong because their logic is bad. They are wrong because they answered late.
Decimas is algorithmic trading software built around that single problem. It is a deterministic trading runtime — a place to observe markets, model what happens next, and route orders inside the time you actually have. The work is infrastructure, not prophecy. It does not promise returns; it promises that the decision you ship is the decision you reasoned about.
A trade is a decision under a clock
The defining constraint of any algorithmic trading platform is not how clever the strategy is. It is how much time passes between a signal and the action it justifies. That interval is a budget, and every component spends from it: market ingest, feature lookup, simulation, the policy check, the order itself.
Decimas starts from that budget and works inward. You declare the time allowed between observing a price and committing an order. Everything downstream is engineered to fit — or it does not ship.
If your software answers after the book has moved, it answered a question nobody asked.
This is why Decimas is snapshot-first. Instead of querying a moving market mid-decision, it reasons over a consistent snapshot of state — the book, positions, and risk as they stood at one instant. The snapshot is the unit of thought, and it is also the unit of record. The market you decided against is the market you can reconstruct later.
The loop
Decimas runs the same four-step loop on every cycle, whether it fires once a minute or thousands of times a second.
- Observe the market — ingest the order book, trades, and positions into a single coherent snapshot.
- Simulate outcomes — run candidate strategies forward against that snapshot, including fills, slippage, and impact.
- Route execution — translate the chosen plan into orders and send them through the venue layer.
- Measure risk — score the resulting position, feed it back, and gate the next cycle.
The loop closes on itself. Risk measured at the end of one cycle becomes a constraint on the simulation at the start of the next, so the system tightens or loosens its own behavior as conditions change.
Simulation is the product, not a sidecar
In most stacks, simulation is something you do beforehand, in a notebook, against last week's data. In Decimas it runs inside the live loop, on the current snapshot, under the same budget as production. The question is never only "what did this strategy do in backtest" but "what will this order do to this book, right now, given what I hold."
That is what makes it a quantitative trading platform rather than a strategy library. Execution strategies, portfolio intelligence, and risk are not separate products bolted together — they are views over one simulated state. A portfolio constraint and an execution tactic argue against the same model of the market, so they cannot quietly disagree.
Machine learning trading algorithms live here too, as policies. A model proposes; the policy layer decides whether the proposal is allowed. Position limits, exposure caps, and venue rules are enforced as the order is formed, not discovered after it fills. A fast wrong order is still a wrong order, so the correctness gates ship alongside the speed.
Speed you can defend
The Mojo-accelerated core exists so the loop fits its budget without thinning out the reasoning inside it. Speed is not the goal; it is the room that lets simulation and risk checks stay rich enough to trust. The point of an AI algorithmic trading system is not to act sooner than everyone else — it is to act on a decision you can stand behind.
Decimas is deterministic by design, and that determinism is what makes replay possible. Given the same snapshot and the same policies, the runtime produces the same orders, every time. When something behaves strangely in production, you do not reason about it from logs and guesswork. You replay the exact cycle and watch the decision form.
For a trading desk, that closes a gap most algorithmic trading software leaves open. Audit stops being a reconstruction and becomes a recording. A compliance question, a risk review, a post-mortem on a bad fill — each resolves to the same artifact: the cycle, re-run, identical.
Where this goes
The harder markets get, the more the advantage moves from having a faster model to having a faster, accountable decision. A quantitative trading platform that can simulate, route, and replay inside one budget turns latency from a tax into a design parameter — something you spend deliberately rather than lose by accident. Decimas is built for the desks that would rather know exactly what they did than guess quickly. As models grow more capable, the constraint that matters most is still the clock, and the discipline is still the same: observe, simulate, route, measure — and be able to prove it.