Machine Learning: LVQ-based Strategy

The Machine Learning: LVQ-based Strategy indicator utilizes a Learning Vector Quantization neural network framework to classify market trends and generate predictive trade signals based on competitive learning.

Usage

The Usage section describes how the script can be used to identify potential market entries and exits through supervised learning classification. Traders can use this tool to determine if specific technical indicators possess predictive value for a given asset.

The script identifies "Buy" and "Sell" opportunities by comparing current market features against trained prototype vectors. When the winning vector indicates a bullish state, a green label appears below the bar; conversely, a bearish state triggers a red label above the bar. Crosses indicate the closure of the current trade.

Signal Filtering

Users can refine signals using the "Filter Signals by" setting:

• Volatility: Only allows signals when short-term ATR exceeds long-term ATR, ensuring trades occur during active market expansion.
• Volume: Filters signals based on a Volume RSI and HMA oscillator to ensure sufficient liquidity.
• Both: Combines both volatility and volume requirements for higher conviction entries.

Details

The LVQ algorithm is a prototype-based supervised learning classification task that trains weights through a competitive, "winner-take-all" approach. Unlike standard backpropagation, LVQ adjusts prototype vectors (weights) in the feature space to better represent specific classes (Buy vs. Sell).

The execution follows these steps:

1. Initialization: Weight vectors for bullish and bearish classes are initialized.
2. Training: During the user-defined "Training Period," the script processes features—specifically RSI, CCI, and ROC—to adjust the weights based on the learning rate and number of epochs.
3. Classification: Once the training period ends, the algorithm uses the Euclidean distance to determine which prototype vector the current market data most closely resembles.
4. Signal Generation: The resulting classification is passed through optional filters to produce final visual signals.

Note: This implementation is a predictive model where signals may recalculate as the neural network adjusts to historical data points during its training phase.

Settings

• Dataset: Defines the source price data (default is Close).
• Lookback Window: Sets the period length for the underlying feature indicators (RSI, CCI, ROC).
• Learning Rate: Controls the magnitude of weight adjustments during each training epoch.
• Epochs: The number of times the training process iterates over the data.
• Filter Signals by: Options to filter outputs by Volatility, Volume, or Both.
• Reverse Signals: Inverts the Buy and Sell logic if the model is yielding contrarian results.
• Training Period (Year/Month/Day): Sets the start and stop dates for the supervised learning phase.
• Lot Size: Used to calculate the cumulative return and performance metrics shown in the info panel.
• Show Information: Toggles the visibility of the performance dashboard (Winrate, Profit, Trade count).

FAQ

How do I adjust the sensitivity of the signals? The sensitivity is primarily controlled by the Learning Rate and the number of Epochs. A higher learning rate makes the model more reactive to recent changes, while more epochs allow the model to converge more tightly on historical patterns.

Why are some signals appearing only after the training period? The algorithm requires a designated historical window to "learn" the relationship between the features (RSI, CCI, ROC) and price movement. Signals generated after the "Training Stop Year" are based on the weights established during that initial phase.

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