Nodes: ML & Execution

The nodes on this page form the final stage of your strategy DAG. They are split into two distinct categories based on how they interact with the rest of the graph:

ML / Reference Nodes (Sources)

Classifier, Regressor, and Strategy Reference nodes are self-contained signal sources. They have no input handles — they do not receive conditions, filters, or logic from other nodes. Instead they produce a signal that flows downstream via a single signal output handle.

Because these nodes are trained or referenced externally, they evaluate independently of the graph. Their output is consumed by Condition nodes (which pivot the categorical or continuous signal into a boolean), and ultimately by Entry and Exit nodes.

Execution Nodes (Terminals)

Entry and Exit nodes are the terminal nodes of the DAG. Every upstream path — condition chains, filters, logic gates, signal gates — eventually converges here. When an Entry or Exit node evaluates to true, the strategy opens or closes a position.

These nodes are the bridge between the abstract strategy graph and real trading activity. They translate boolean logic into actual order instructions that the backtester (and later, the live execution engine) act on.

A Classifier node wraps a trained machine-learning classification model that predicts market direction. During backtesting, the node loads the serialized model, feeds it the current feature vector, and emits a categorical prediction — Short, Neutral, or Long — on every bar.

Inputs (self-contained)

This node has no input handles. It reads features from the market data pipeline internally.

Parameters

Model Info * required

Dropdown listing every trained classifier available from your ML pipeline. Each entry shows:

• Model name — the label you assigned during training
• Training symbol / timeframe — e.g., BTCUSDT / 1h
• Accuracy — overall classification accuracy on the validation set
• Precision / Recall / F1 — per-class metrics where available

The model must have id != 0 (i.e., it must be a real persisted model, not a placeholder entry). If you delete a model that a strategy references, the strategy will fail validation and refuse to run until you reassign.

Alias * required

A unique, machine-friendly identifier used to reference this node elsewhere in the DAG (primarily by Condition nodes that consume its signal output). Must be unique across the entire strategy. Use lowercase snake_case: ml_classifier_btc, not ML Classifier BTC.

Display Name optional

A human-readable label shown on the node in the canvas. Defaults to the alias if left blank.

Has Probabilities display only

Read-only indicator. If the underlying model was trained with probability output (e.g., a RandomForest or XGBoost classifier with predict_proba support), this toggle is on and the node exposes confidence scores alongside each prediction. Models without probability support (like certain SVM configurations) show this as off.

Confidence Margin

• Type: float
• Range: 0.0 – 1.0
• Default: 0.5

The minimum prediction confidence required for the node to emit a directional signal. The model assigns a probability to each class (Short / Neutral / Long). Only if the winning class exceeds this threshold does the signal fire. Otherwise, the output is Neutral (no conviction).

Examples:

• 0.5 — any above-even prediction fires (default, most permissive)
• 0.7 — the model must be at least 70% confident in its prediction. Filters out weak signals and reduces entries significantly.
• 0.9 — only the model's highest-conviction calls. Very few entries; use with a broad "or" entry combination mode.

Direction Bias optional

Restricts which directional signals the node emits. Useful when a classifier is trained on both directions but you only want it to influence one side of your strategy.

• all (default) — emit Short, Neutral, and Long as predicted
• long_only — only emit Long and Neutral (Short predictions are suppressed to Neutral)
• short_only — only emit Short and Neutral (Long predictions suppressed)
• long_short — emit Long and Short, but suppress Neutral (forces a directional call on every bar)

Output Handle

A Regressor node wraps a trained ML regression model that predicts a continuous value — typically the expected percentage return over a forward window. Unlike the Classifier, which emits categories, the Regressor emits a numeric value that you can compare with percentage-based thresholds in a downstream Condition node.

Inputs (self-contained)

No input handles. The node reads features from the market data pipeline internally.

Parameters

Model Info * required

Dropdown of trained regressor models from your ML pipeline. Each entry shows:

• Model name — training label
• Training symbol / timeframe
• R-squared, MAE, RMSE — regression quality metrics
• Prediction horizon — how many bars forward the model was trained to predict (e.g., "next 4 bars")

The model must have id != 0. Selecting a model trained on a different symbol or timeframe than your strategy operates on will produce meaningless predictions — the system warns but does not block this.

Alias * required

Unique identifier for the node. Example: ml_regressor_eth.

Display Name optional

Human-readable label for the canvas node. Defaults to alias.

Output Handle

Typical Usage Patterns

A Strategy Reference node lets one strategy consume the signals of another saved strategy. This enables strategy composition — building higher-level strategies from lower-level building blocks — without duplicating the underlying DAG.

Inputs (self-contained)

No input handles. The referenced strategy runs its own DAG internally; this node only surfaces its entry/exit signals as output.

Parameters

Strategy ID * required

Dropdown of all saved strategies in your account. Each entry displays:

• Strategy name
• Symbol — e.g., BTCUSDT, ETHUSDT
• Timeframe — e.g., 15m, 1h, 4h, 1d
• Status tag — Draft / Backtested / Live

The symbol and timeframe of the referenced strategy must match the referencing strategy. If they differ, validation fails with a clear error message. You cannot reference a strategy that references this strategy (circular dependency detection).

Alias * required

Unique identifier. Example: ref_trend_detection.

Display Name optional

Human-readable canvas label.

Direction Bias optional

Filters which signal types pass through from the referenced strategy. Same four options as the Classifier node:

• all — all signal types pass through
• long_only — only Long Entry and Long Exit signals
• short_only — only Short Entry and Short Exit signals
• long_short — all entry/exit signals; suppress any neutral output

Output Handle

Use Case: Strategy Composition

A classic pattern is separating signal generation from trade management:

1. Base strategy ("Trend Detection") — a DAG that detects trend direction using EMA crossovers, ADX, and volume confirmation. It emits Long Entry and Short Entry signals.
2. Composite strategy ("Trend Following") — references the base strategy via a Strategy Reference node, then layers on additional filters (e.g., a Volatility Filter to avoid entry during high-VIX regimes) and risk management exits.

This means you can backtest and refine the "Trend Detection" core independently, then reuse it across multiple composite strategies without copy-pasting the DAG.

The Entry node is the terminal node where your entire upstream signal chain converges into a trading decision. If the DAG is a circuit, the Entry node is the switch that closes and sends current to the order engine. Getting this node right is the difference between a strategy that trades and one that sits idle.

Direction

* required — cannot be changed after creation

Each Entry node handles exactly one direction. To trade both long and short in the same strategy, create two Entry nodes — one Long, one Short. They are visually distinct on the canvas:

• Long (green node header) — enters a buy / long position. Executes a market buy order in the backtester.
• Short (red node header) — enters a sell / short position. Executes a market sell order in the backtester.

Direction is immutable after creation. If you need to switch directions, delete the node and create a new one.

Handle Layout

The Entry node has two sets of input handles on different edges:

Condition Inputs (LEFT edge)

Five handles labeled cond-0 through cond-4, stacked vertically between 20% and 78% of the node height. These accept:

• Condition node result output — a boolean true/false based on the condition's operator and threshold
• Logic node out output — the result of an AND/OR/NOT/XOR logic gate combining multiple upstream conditions
• Signal Gate node out output — a gated signal that only passes through when enabled

Connecting a wire to any condition slot auto-assigns it to the first available free slot. You can drag wires to a specific slot by dropping on the exact handle position.

Filter Inputs (TOP edge)

Five handles labeled filter-0 through filter-4, stacked horizontally between 10% and 90% of the node width. These accept:

• Filter node gate output — a boolean that acts as a permission check

Critical rule: Unlike conditions (which combine according to the combination mode), ALL active filters must pass (true) for the entry to fire. Filters are AND-ed together unconditionally. A single active filter returning false blocks the entry regardless of conditions.

Combination Mode * required

Determines how the connected condition inputs are logically combined to produce a single entry decision. Each mode has its own configuration panel that appears when selected.

Signal Processing optional

Post-processing rules applied to the entry decision after the combination logic evaluates. These control how repeated or opposite signals are handled when a position is already open.

Common Anti-Patterns

A Signal Exit node closes positions based on conditions evaluated on each bar — the DAG-driven counterpart to the risk-management exit. While the Risk Management Exit operates on price levels (stop loss, take profit), the Signal Exit operates on indicator values and model signals.

Role: exit

This is a signal-based exit node. It evaluates its connected conditions on every bar while a position is open and closes the position (via market order) when the conditions are met.

Applies To * required

• long — only closes long positions
• short — only closes short positions
• both — closes any open position regardless of direction

If you have separate exit logic for long and short positions, create two Signal Exit nodes — one for each direction.

Condition Inputs (LEFT edge)

Five handles labeled cond-0 through cond-4 on the left edge, identical layout to the Entry node's condition inputs. These accept Condition result, Logic out, and Signal Gate out outputs.

Combination Logic

Signal Exit nodes support a subset of combination modes (no not or sequential, as these patterns don't apply to exit logic):

• and — ALL connected exit conditions must be true to close the position
• or — ANY connected exit condition triggers the close (most common)
• weighted — weighted sum of true exit conditions exceeds threshold
• majority — majority of exit conditions are true

Common Exit Condition Examples

A Risk Management Exit node is self-contained — it has no input handles. It is automatically applied to every trade the strategy opens. Unlike the Signal Exit (which queries the DAG on each bar), the Risk Management Exit monitors price relative to entry and closes the position when predefined thresholds are breached.

Role: risk-management

Self-contained, no handles. Every trade opened by the strategy is automatically subject to these risk parameters. You cannot opt a trade out of risk management — if you attach this node to the strategy, it applies to all positions.

Applies To * required

• long — risk parameters apply only to long positions
• short — risk parameters apply only to short positions
• both — same risk parameters apply to all positions

If long and short positions need different risk profiles (e.g., wider stop for shorts in an uptrending market), create two Risk Management Exit nodes.

Stop Loss (SL)

Automatically closes the position when price moves against you by a specified amount. Always denominated from the entry price.

Take Profit (TP)

Automatically closes the position when price moves in your favor by a specified amount. Locks in gains without relying on a signal-based exit.

Trailing Stop Loss (TSL)

A dynamic stop loss that moves with the price as it moves in your favor, but stays fixed when price reverses. This lets winners run while protecting accumulated gains.

Time Stop

Closes the position after a fixed number of bars, regardless of P&L. Prevents capital from being tied up in stagnant positions and enforces a maximum trade duration.