How QuantiX IQ Works
This page explains what actually happens behind the scenes when you describe a trading strategy to QuantiX IQ in chat — the pipeline of specialized AI agents that turns your written request into a fully trained, backtested, audited strategy. See QuantiX IQ for a product-level introduction; this page goes one level deeper, into how the system is built and how it makes decisions.
The chat session
Every conversation with QuantiX IQ is a thread. Each message you send starts a run — one full (or partial) pass through the agent pipeline described below. While a run is in progress, the chat interface streams each agent's progress to you live, in order, as a series of stage cards: what that agent decided, the parameters it chose, and (once training/backtesting happens) the resulting metrics. You can return to a thread later — its full history of runs and agent outputs is reloaded exactly as it was streamed the first time.
If a run is taking too long or heading in a direction you don't want, it can be stopped mid-flight; the pipeline halts at the next safe checkpoint rather than continuing to completion.
Before anything else: is this a strategy request?
The very first thing that happens to your message is a lightweight relevance check. If what you wrote is a greeting, small talk, or unrelated to building/discussing a trading strategy, QuantiX IQ replies conversationally (and, if it looks like you were trying to describe a strategy, prompts you for the details it needs) instead of running the full agent pipeline. Only requests that are recognizably about building, running, backtesting, or discussing a trading strategy proceed further.
The agent pipeline
Once your request passes that check, it flows through a fixed sequence of specialized agents. Each agent has one job, reads the state the previous agents have built up so far, and hands off to the next one. The high-level flow looks like this:
1. Strategy Architect
Reads your written request and works out the shape of the strategy you're asking for: does it need an AI-trained Long entry, an AI-trained Short entry, a Long exit, a Short exit, and/or entries or exits you described explicitly yourself (a Manual rule, e.g. "enter when RSI crosses above 30")? Any combination of these can be requested in a single message — the Strategy Architect's decision here determines exactly which of the branches below actually run for this request.
2. Asset Agent
Determines the trading universe and time boundaries for the strategy: which pair(s) to trade, the main (base) timeframe, any supplementary timeframes to pull additional indicators from, and — critically — the training date range versus the backtesting date range. These two ranges are always kept separate and non-overlapping, so the strategy is never backtested on the same data it was trained on.
3. The Long and Short branches
If the Strategy Architect asked for an AI-trained Long and/or Short signal, that side runs through its own full sub-pipeline (the Long and Short branches are structurally identical, just optimizing in opposite directions, and run one after the other rather than at the same time):
Step | What it does |
|---|---|
Target Agent | Defines what "success" means for the model to learn — typically a stop-loss/take-profit classification target, with SL/TP levels computed dynamically from the asset's own recent volatility (ATR) rather than fixed percentages. |
Indicator Agent | Selects the technical indicators and features the model will train on — moving averages, oscillators (RSI, MACD), volatility bands, pivot/Fibonacci levels, and more, computed across whichever timeframes the Asset Agent configured. |
Model Selection | Chooses the machine learning model architecture for this signal. |
Hyperparameter Tuning | Tunes the chosen model's hyperparameters (e.g. XGBoost settings), including rebalancing class weights so the model isn't overwhelmed by the natural imbalance between "target hit" and "target missed" outcomes. |
Training | Trains the model on the Target Agent's labels and the Indicator Agent's features, using only the Train Time Range from the Asset Agent. |
Signal Entry / Signal Exit | Turns the trained model's predictions into concrete entry and exit signal rules. |
4. Manual entries and exits
If the Strategy Architect instead recognized an explicit rule in your request (or in addition to an AI branch), the corresponding entry/exit is built directly from the logic you described, without training a model.
5. Signal Composer & Signal Analyzer — the first quality gate
Once every requested branch (Long, Short, and/or Manual) has produced its signals, the Signal Composer merges them into one coherent strategy definition. The Signal Analyzer then audits that definition before any backtest is run, checking things like:
Overfitting: how much the model's performance drops from training data to held-out data (both ROC-AUC and Precision-Recall gaps are checked, and must stay small).
Signal precision: whether the model's positive predictions are actually reliable, not just better than random chance by a trivial margin.
Gain-to-pain ratio: whether the signal's expected reward-to-risk profile clears a minimum bar.
If any of these checks fail, the Signal Analyzer doesn't just reject the strategy — it routes control back to whichever earlier stage is most likely responsible (re-running the Indicator Agent with different features, re-tuning hyperparameters, retraining, or rebuilding the composed signal), and the pipeline tries again. This retry loop is capped at a fixed number of attempts so a stubborn strategy configuration can't loop forever; if it can't pass after enough attempts, the run ends and reports what went wrong.
6. Pyramiding & Stop-Loss/Take-Profit
Once the underlying signals pass the quality gate, the Pyramiding Agent configures money management and position-sizing behavior (whether and how the strategy can add to a position while it's open, and any caps on exposure), and the TP/SL Agent finalizes the strategy's actual stop-loss/take-profit and maximum-hold-time execution rules.
7. Backtest Executor
Runs a full historical backtest of the finished strategy over the Backtest Time Range the Asset Agent set aside — the period the strategy was never trained or tuned on.
8. Judge Agent — the final quality gate
Audits the backtest results against strict, fixed production criteria before the strategy is considered ready to use: minimum expected average return, minimum Sharpe ratio, a floor on worst expected drawdown, and a ceiling on how long the strategy can take to recover from a drawdown. If the backtest doesn't clear these bars, the Judge Agent routes control back to the Pyramiding Agent to try a different money-management configuration and re-backtest — again, up to a capped number of attempts before the run ends and reports the shortfall.
What you get at the end
A strategy that passes both quality gates is a complete, trained, backtested strategy — the same kind of object you'd get from converting a backtest report into a strategy manually (see What is a Strategy?), ready to review, publish, or run as a bot, with every stage of how it was built visible in the chat history.