Machine Learning: Perceptron-based strategy

The Machine Learning: Perceptron-based strategy indicator utilizes a single-layer artificial neural network to classify price action and generate trend-following signals based on learned data patterns. It aims to identify the relationship between historical inputs and current price direction to provide actionable buy and sell entries.

Usage

The Usage section describes how the script can be used to identify market trends. The indicator functions by retraining a perceptron on every new bar, attempting to classify whether the current price action is bullish or bearish based on historical sample data.

• Signal Generation: Buy signals are triggered when the perceptron classifies the price action as positive (outcome > 0) while the DMI is above the specified threshold. Sell signals occur when the classification is negative and the DMI is below the threshold.
• Filtering: Users can apply Volatility or Volume filters to reduce noise. For example, selecting "Volatility" ensures signals only trigger when market range (ATR) is expanding.
• Holding Period: The script includes a built-in exit mechanism based on a "Holding Period." Once the specified number of bars has passed since a signal, a stop cross is plotted, indicating the trade should be closed.
• Repainting Note: Because the perceptron retrains and updates its weights dynamically on each bar, signals may change their appearance or position in a live environment. Using the TradingView playback feature is recommended to observe this behavior.

Details

The script implements a Learning Perceptron, the simplest form of an artificial neural network (ANN). It consists of a single neuron that processes inputs, applies weights and a bias, and passes the result through a step function activation: $f(z) = 1$ if $z > 0$, else $-1$.

The training process involves:

1. Initialization: Weights and bias are initially set.
2. Learning Rule: The script iterates through a "Sample Size" of historical bars (Epochs) to update weights based on the error of each observation.
3. Optimization: The weight parameters are adjusted to move the classifier’s decision boundary toward a correct classification of the last observation.
4. Convergence: Through repeated sweeps (epochs), the weights converge to a state that best models the recent price relationship.

Settings

General Settings

• Time Frame: Defines the timeframe from which the data is pulled for calculation.
• Sample Size: The number of historical bars used for training the perceptron.
• Period: The lookback period used for the internal Moving Average and DMI calculations.
• Threshold [0..100]: Sets the sensitivity level for the DMI filter.
• Filter: Choose between Volatility, Volume, Both, or None to qualify signals.
• Holding Period: The maximum number of bars to stay in a trade before a stop signal is generated.

Visual Settings

• Ma: Toggles the visibility of the internal Moving Average curve.
• Bar Color: Colors the price bars based on the current detected trend.
• Trailing: Displays a line based on linear regression and MG calculation to assist in trailing stops.
• Info: Displays a label on the chart showing current DMI values and the sum of errors from the perceptron training.

FAQ

How does the perceptron "learn" in this script?

The script uses a for-loop to iterate through historical data points (epochs). During each iteration, it calculates the prediction error and updates the weights ($weight1$, $weight2$) and the bias to minimize that error.

Why do the signals repaint?

The signals repaint because the neural network's weights are recalculated on every new bar using the most recent data. As the training set shifts forward, the optimized boundary for previous bars may change.

How do I access this indicator?

You can get access on the LuxAlgo Library for charting platforms like TradingView, MetaTrader (MT4/MT5), and NinjaTrader for free.