Deutsch DE
Auf GitHub ansehen

Mnist Pattern Classifier Strategy

Origin

The strategy is a StockSharp port of the MetaTrader 5 expert TestMnistOnnx.mq5 (MQL ID 47225). The original script exposes an interactive canvas where the user draws digits that are classified by a bundled MNIST ONNX model. The StockSharp version keeps the spirit of pattern recognition but replaces the hand-drawn canvas with a rolling matrix built from finished candles.

Concept

  1. A rolling window of LookbackPeriod completed candles (default 28) is treated as a 28×28 grid similar to a MNIST image.
  2. Several statistical features – range compression, trend strength, momentum, RSI deviation and ATR normalization – are blended into a synthetic "confidence" score that mimics the neural network probability produced by the MQL expert.
  3. The resulting features are mapped to one of ten pattern classes (0–9). Each class represents a market regime (flat, trend, breakout, pullback, reversal, etc.).
  4. When the detected class matches the user-selected TargetClass and the synthetic confidence is above ConfidenceThreshold, the strategy opens or reverses a position in the indicated direction. Positions are flattened if the class changes or the confidence falls below the threshold.

Parameters

Parameter Default Description
LookbackPeriod 28 Number of finished candles that are converted into the MNIST-like grid.
TargetClass 1 Class index (0–9) that should trigger trading actions.
ConfidenceThreshold 0.6 Minimum synthetic probability that allows order submission.
Volume 1 Order volume for new positions.
CandleType 5-minute timeframe Data type subscribed for candle updates.

Pattern Classes

Class Meaning
0 Flat or low-volatility consolidation.
1 Sustained bullish trend.
2 Sustained bearish trend.
3 Breakout to the upside with strong follow-through.
4 Breakout to the downside with strong follow-through.
5 Wide volatile range without clear bias.
6 Bullish pullback inside an uptrend.
7 Bearish pullback inside a downtrend.
8 Bullish reversal after an extended decline.
9 Bearish reversal after an extended advance.

Trading Rules

  • Trades only on completed candles to remain in sync with the original expert that reacted to finalized drawings.
  • Uses market orders (BuyMarket, SellMarket) and flattens before reversing to mimic the single-position behaviour of the original script.
  • Confidence scaling is bounded to [0, 1]. Raising ConfidenceThreshold filters out weaker signals.
  • The strategy does not manage protective stops; risk management is expected to be configured externally in StockSharp.

Usage Tips

  • Select a candle type that reflects the market rhythm you want to analyse. Shorter timeframes react faster but are noisier.
  • Optimise TargetClass and ConfidenceThreshold together – some classes are naturally rarer and may require lower thresholds.
  • The synthetic pattern classifier is deterministic; there is no dependency on external ONNX runtime libraries.
  • Combine with the built-in risk protection tools available in StockSharp (such as StartProtection) to control exposure.

Differences from the Original

  • Interactive drawing and ONNX inference are replaced by fully automated candle analysis.
  • The "confidence" is a deterministic blend of indicators rather than a neural network probability.
  • Trading logic is added to convert pattern recognition into actionable orders.
  • The MNIST resource file is not required in the StockSharp environment.
using System;
using System.Linq;
using System.Collections.Generic;

using Ecng.Common;
using Ecng.Collections;
using Ecng.Serialization;

using StockSharp.Algo.Indicators;
using StockSharp.Algo.Strategies;
using StockSharp.BusinessEntities;
using StockSharp.Messages;

namespace StockSharp.Samples.Strategies;

/// <summary>
/// Port of the MetaTrader expert TestMnistOnnx.
/// Converts a rolling grid of candle closes into pattern classes and trades on a selected class.
/// The original mouse-drawn image is replaced with market data derived features.
/// </summary>
public class MnistPatternClassifierStrategy : Strategy
{
	private readonly StrategyParam<int> _lookbackPeriod;
	private readonly StrategyParam<int> _targetClass;
	private readonly StrategyParam<decimal> _confidenceThreshold;
	private readonly StrategyParam<DataType> _candleType;

	private RelativeStrengthIndex _rsi = null!;
	private AverageTrueRange _atr = null!;

	private readonly Queue<decimal> _closeWindow = new();
	private decimal _firstClose;
	private decimal _previousClose;

	private int _lastClass = -1;
	private decimal _lastConfidence;
	private int _cooldown;

	private enum PatternBiases
	{
		Neutral,
		Bullish,
		Bearish,
	}

	/// <summary>
	/// Number of finished candles that form the MNIST-like grid.
	/// </summary>
	public int LookbackPeriod
	{
		get => _lookbackPeriod.Value;
		set => _lookbackPeriod.Value = value;
	}

	/// <summary>
	/// Pattern class (0-9) that will trigger trading actions.
	/// </summary>
	public int TargetClass
	{
		get => _targetClass.Value;
		set => _targetClass.Value = value;
	}

	/// <summary>
	/// Minimum confidence required before orders are sent.
	/// </summary>
	public decimal ConfidenceThreshold
	{
		get => _confidenceThreshold.Value;
		set => _confidenceThreshold.Value = value;
	}


	/// <summary>
	/// Candle type that feeds the pattern grid.
	/// </summary>
	public DataType CandleType
	{
		get => _candleType.Value;
		set => _candleType.Value = value;
	}

	/// <summary>
	/// Initializes a new instance of the <see cref="MnistPatternClassifierStrategy"/> class.
	/// </summary>
	public MnistPatternClassifierStrategy()
	{
		_lookbackPeriod = Param(nameof(LookbackPeriod), 14)
		.SetRange(10, 200)

		.SetDisplay("Lookback", "Number of candles converted into the pattern grid", "Pattern");

		_targetClass = Param(nameof(TargetClass), 1)
		.SetRange(0, 9)
		.SetDisplay("Target Class", "Pattern class that should be traded", "Pattern");

		_confidenceThreshold = Param(nameof(ConfidenceThreshold), 0.2m)
		.SetRange(0m, 1m)
		.SetDisplay("Confidence", "Minimum classification confidence", "Pattern");


		_candleType = Param(nameof(CandleType), TimeSpan.FromMinutes(5).TimeFrame())
		.SetDisplay("Candle Type", "Primary timeframe used for the pattern", "General");
	}

	/// <inheritdoc />
	public override IEnumerable<(Security sec, DataType dt)> GetWorkingSecurities()
	{
		yield return (Security, CandleType);
	}

	/// <inheritdoc />
	protected override void OnReseted()
	{
		base.OnReseted();

		_closeWindow.Clear();
		_firstClose = 0m;
		_previousClose = 0m;
		_lastClass = -1;
		_lastConfidence = 0m;
		_cooldown = 0;
	}

	/// <inheritdoc />
	protected override void OnStarted2(DateTime time)
	{
		base.OnStarted2(time);

		StartProtection(
			takeProfit: new Unit(3, UnitTypes.Percent),
			stopLoss: new Unit(2, UnitTypes.Percent));

		_rsi = new RelativeStrengthIndex
		{
			Length = LookbackPeriod,
		};

		_atr = new AverageTrueRange
		{
			Length = LookbackPeriod,
		};

		var subscription = SubscribeCandles(CandleType);
		subscription
		.Bind(_rsi, _atr, ProcessCandle)
		.Start();
	}

	private void ProcessCandle(ICandleMessage candle, decimal rsiValue, decimal atrValue)
	{
		if (candle.State != CandleStates.Finished)
		return;

		UpdateWindow(candle.ClosePrice);

		if (_closeWindow.Count < LookbackPeriod)
		{
			_previousClose = candle.ClosePrice;
			return;
		}

		if (_cooldown > 0)
		{
			_cooldown--;
			_previousClose = candle.ClosePrice;
			return;
		}

		var pattern = ClassifyPattern(candle.ClosePrice, rsiValue, atrValue);

		_lastClass = pattern.PatternClass;
		_lastConfidence = pattern.Confidence;

		if (pattern.Confidence >= ConfidenceThreshold && pattern.Bias != PatternBiases.Neutral && Position == 0)
		{
			ExecuteBias(pattern.Bias);
		}

		_previousClose = candle.ClosePrice;
	}

	private void ExecuteBias(PatternBiases bias)
	{
		switch (bias)
		{
		case PatternBiases.Bullish:
			if (Position == 0)
			{
				BuyMarket();
				_cooldown = 50;
			}

			break;
		case PatternBiases.Bearish:
			if (Position == 0)
			{
				SellMarket();
				_cooldown = 50;
			}

			break;
		default:
			break;
		}
	}

	private void UpdateWindow(decimal close)
	{
		_closeWindow.Enqueue(close);

		if (_closeWindow.Count > LookbackPeriod)
		{
			_closeWindow.Dequeue();
		}

		_firstClose = _closeWindow.Count > 0 ? _closeWindow.Peek() : 0m;
	}

	private PatternResult ClassifyPattern(decimal currentClose, decimal rsiValue, decimal atrValue)
	{
		var stats = CalculateStatistics(currentClose, rsiValue, atrValue);

		var trendStrength = stats.TrendStrength;
		var rangeStrength = stats.RangeStrength;
		var breakoutRange = stats.BreakoutThreshold;
		var rangePosition = stats.RangePosition;
		var momentum = stats.Momentum;
		var rsi = stats.Rsi;
		var atr = stats.AtrNormalized;

		// Compute a blended confidence score similar to the ONNX output probability.
		var confidence = Math.Min(1m, (trendStrength + rangeStrength + Math.Min(1m, Math.Abs(momentum) / stats.MomentumThreshold) + stats.RsiDeviation + atr) / 5m);

		if (rangeStrength < stats.FlatThreshold)
		{
			return new PatternResult(0, Math.Max(confidence, 0.4m), PatternBiases.Neutral);
		}

		if (trendStrength >= stats.TrendThreshold)
		{
			if (rangePosition >= 0.75m && rangeStrength >= breakoutRange)
			{
				return new PatternResult(3, confidence, PatternBiases.Bullish);
			}

			if (momentum < 0m)
			{
				return new PatternResult(6, confidence * 0.8m, PatternBiases.Bullish);
			}

			return new PatternResult(1, confidence, PatternBiases.Bullish);
		}

		if (trendStrength <= -stats.TrendThreshold)
		{
			if (rangePosition <= 0.25m && rangeStrength >= breakoutRange)
			{
				return new PatternResult(4, confidence, PatternBiases.Bearish);
			}

			if (momentum > 0m)
			{
				return new PatternResult(7, confidence * 0.8m, PatternBiases.Bearish);
			}

			return new PatternResult(2, confidence, PatternBiases.Bearish);
		}

		if (rangeStrength >= breakoutRange)
		{
			return new PatternResult(5, confidence * 0.9m, PatternBiases.Neutral);
		}

		if (rangePosition <= 0.4m && rsi >= 55m)
		{
			return new PatternResult(8, confidence * 0.85m, PatternBiases.Bullish);
		}

		if (rangePosition >= 0.6m && rsi <= 45m)
		{
			return new PatternResult(9, confidence * 0.85m, PatternBiases.Bearish);
		}

		return new PatternResult(0, confidence * 0.7m, PatternBiases.Neutral);
	}

	private PatternStatistics CalculateStatistics(decimal currentClose, decimal rsiValue, decimal atrValue)
	{
		var window = _closeWindow.ToArray();
		decimal min = decimal.MaxValue;
		decimal max = decimal.MinValue;

		foreach (var value in window)
		{
			if (value < min)
			min = value;

			if (value > max)
			max = value;
		}

		var first = _firstClose;
		var last = currentClose;
		var range = max - min;
		var rangeStrength = first != 0m ? range / first : 0m;
		var trend = first != 0m ? (last - first) / first : 0m;
		var momentum = _previousClose != 0m ? (last - _previousClose) / _previousClose : 0m;
		var rsiDeviation = Math.Min(1m, Math.Abs(rsiValue - 50m) / 50m);
		var atrNormalized = first != 0m ? Math.Min(1m, atrValue / first) : 0m;

		var rangePosition = range > 0m ? (last - min) / range : 0.5m;

		const decimal baseThreshold = 0.001m;
		var trendThreshold = baseThreshold;
		var breakoutThreshold = baseThreshold * 1.4m;
		var flatThreshold = baseThreshold * 0.3m;
		var momentumThreshold = baseThreshold;

		return new PatternStatistics
		{
			TrendStrength = trend,
			RangeStrength = rangeStrength,
			BreakoutThreshold = breakoutThreshold,
			FlatThreshold = flatThreshold,
			RangePosition = rangePosition,
			Momentum = momentum,
			Rsi = rsiValue,
			AtrNormalized = atrNormalized,
			TrendThreshold = trendThreshold,
			MomentumThreshold = momentumThreshold,
			RsiDeviation = rsiDeviation,
		};
	}

	private readonly struct PatternResult
	{
		public PatternResult(int patternClass, decimal confidence, PatternBiases bias)
		{
			PatternClass = patternClass;
			Confidence = confidence;
			Bias = bias;
		}

		public int PatternClass { get; }

		public decimal Confidence { get; }

		public PatternBiases Bias { get; }
	}

	private readonly struct PatternStatistics
	{
		public decimal TrendStrength { get; init; }
		public decimal RangeStrength { get; init; }
		public decimal BreakoutThreshold { get; init; }
		public decimal FlatThreshold { get; init; }
		public decimal RangePosition { get; init; }
		public decimal Momentum { get; init; }
		public decimal Rsi { get; init; }
		public decimal AtrNormalized { get; init; }
		public decimal TrendThreshold { get; init; }
		public decimal MomentumThreshold { get; init; }
		public decimal RsiDeviation { get; init; }
	}
}