Each day, the neural network outputs a value between 0 and 1. This is converted into a directional signal:
โข Output > 0.5 โ GO LONG โ buy the asset, profit if price rises
โข Output < 0.5 โ GO SHORT โ sell/short the asset, profit if price falls
The position is held for the forecast horizon (1โ7 days), then closed. The strategy uses 80% position sizing โ only 80% of the portfolio is committed to each trade, keeping 20% as a cash buffer.
For multi-day horizons, the daily return is divided by the number of days to avoid over-leveraging on longer predictions.
If the model correctly predicts direction, actual_move is positive (we earn the market move). If wrong, actual_move is negative (we lose).
Trading cost is deducted on every trade as a round-trip cost (open + close). Set this to match your broker/exchange:
โข Crypto CEX (Binance): ~10โ20 bp
โข US Stocks (IBKR): ~1โ5 bp
โข Forex: ~2โ8 bp (spread-dependent)
โข Crypto DEX (Uniswap): ~30โ100 bp
๐ฌ How Backtesting Works
1. Expanding Window Start with 50% of data for training, predict the next day. Then add one more day to training, predict again. Repeat until all data is used. This simulates real trading โ the model never sees future data.
2. Walk-Forward Retraining The model is retrained ~30 times as the training window grows, using warm-starting (previous weights as starting point). This lets the model adapt to changing market conditions over time.
3. Metrics Reported Strategy Return: total P&L after costs. Hit Rate: % of correct directional predictions. Buy & Hold: what simple holding would have earned โ the benchmark to beat.
๐ง Neural Network Architecture
Select architecture. LSTM adds recurrent memory for temporal pattern recognition.
Feed-Forward Neural Network โ SmallData flows in one direction: inputs โ hidden layers โ output. Each neuron receives all outputs from the previous layer, applies a weighted sum, and passes through a sigmoid activation function. With only 12 neurons, this model learns broad trends quickly but may miss subtle patterns. Best when data is limited or you want a fast baseline.
๐ข SIMPLE
Feed-Forward 2 layers (8โ4) ~2s ยท Quick scans
Feed-Forward Neural Network โ MediumSame one-directional architecture, but with 3 hidden layers and 32 neurons total. Uses Leaky ReLU activation, which avoids the "dying neuron" problem of sigmoid โ neurons stay responsive even for weak signals. More layers let the network learn hierarchical patterns: layer 1 detects simple price moves, layer 2 combines them into patterns, layer 3 forms trading signals.
๐ก BALANCED
Feed-Forward 3 layers (16โ10โ6) ~5s ยท Daily use
Feed-Forward Neural Network โ Large4 hidden layers with 70 neurons create a deep network capable of learning complex non-linear relationships between features. The progressive narrowing (32โ20โ12โ6) acts as an information funnel โ compressing many inputs into increasingly abstract representations. Risk: with limited data, deep networks can overfit (memorize noise instead of learning patterns).
๐ COMPLEX
Feed-Forward 4 layers (32โ20โ12โ6) ~12s ยท Deep
LSTM โ Long Short-Term MemoryUnlike feed-forward networks, LSTM has memory cells that remember information across time steps. Each cell has 3 gates: forget (what to discard), input (what new info to store), and output (what to pass forward). This lets it learn temporal sequences โ e.g., "a 3-day RSI decline followed by a volume spike often precedes a reversal." Best for time-series data where order and history matter.
๐ฃ LSTM
Recurrent NN Memory cells ~15s ยท Temporal
16
Input Features
3
Hidden Layers
32
Total Neurons
Feed-Forward
Network Type
1000
Iterations
0.02
Learning Rate
LeakyReLU
Activation
Analyzing assets...0%
๐ Analysis Results
Asset
Price
Signal
Predicted ฮ
Confidence
RSI
MACD
BB Pos
Status
Hit Rate
๐ฌ Backtest Results โ
Actual Price Correct forecast ray Wrong forecast ray First training set boundary
๐ Grid Search โ
Strategy Return (%) for each Architecture ร History combination. โ ๐ Best return ยท โ ๐ฏ Best hit ยท Click ๐พ Save on any cell to save the strategy.
๐ Selected Asset Model โ trains & tests on this asset only (matches โก Analyze Selected)
Architecture
30 days
60 days
90 days
120 days
๐ Cross-Asset Model โ trains on ALL active assets, tests on selected (matches ๐ Analyze All)
Architecture
30 days
60 days
90 days
120 days
๐พ Saved Strategies
No saved strategies yet. Click any grid search cell to save it.
๐ Manage Asset Watchlist
โ
Selected: 0 assets ยท Crypto, forex, commodities, and stocks โ all manageable
๐ง Feature Builder โ Neural Network Inputs
โ
Choose which indicators feed the neural network. "Target Asset" = the asset selected in the main dropdown (changes automatically). Add other assets below for cross-market signals.
Total input features: 16 โ Neural network input layer will have this many nodes
[SYSTEM] NeuroTrader Pro v3 initialized โ LSTM engine ready
๐ Interactive Candlestick Chart
RSI (14)
MACD (12, 26, 9)
๐ง Neural Network Visualizer
Arch:16โ10โ6
Type:Feed-Forward
Connections:0
Status:Idle
Node brightness = activation strength ยท Edge color = weight polarity ยท LSTM nodes show recurrent loops