Enter Platform
Resources
Back to Notebooks
Open in Colab
Signals·Machine Learning Models·Advanced

LSTM Model

Build and train an LSTM neural network for financial time series prediction with properly formatted input sequences, bidirectional layers, dropout regularization, and strict walk-forward validation to avoid information leakage across time periods.

machine-learningtrading-signals

Signals — LSTM for Price Prediction

1. Dependency Installation

[ ]
!pip install pandas numpy plotly scikit-learn tensorflow
Requirement already satisfied: pandas in /usr/local/lib/python3.12/dist-packages (2.2.2)
Requirement already satisfied: numpy in /usr/local/lib/python3.12/dist-packages (2.0.2)
Requirement already satisfied: plotly in /usr/local/lib/python3.12/dist-packages (5.24.1)
Requirement already satisfied: scikit-learn in /usr/local/lib/python3.12/dist-packages (1.6.1)
Requirement already satisfied: tensorflow in /usr/local/lib/python3.12/dist-packages (2.20.0)
Requirement already satisfied: python-dateutil>=2.8.2 in /usr/local/lib/python3.12/dist-packages (from pandas) (2.9.0.post0)
Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.12/dist-packages (from pandas) (2025.2)
Requirement already satisfied: tzdata>=2022.7 in /usr/local/lib/python3.12/dist-packages (from pandas) (2026.1)
Requirement already satisfied: tenacity>=6.2.0 in /usr/local/lib/python3.12/dist-packages (from plotly) (9.1.4)
Requirement already satisfied: packaging in /usr/local/lib/python3.12/dist-packages (from plotly) (26.1)
Requirement already satisfied: scipy>=1.6.0 in /usr/local/lib/python3.12/dist-packages (from scikit-learn) (1.16.3)
Requirement already satisfied: joblib>=1.2.0 in /usr/local/lib/python3.12/dist-packages (from scikit-learn) (1.5.3)
Requirement already satisfied: threadpoolctl>=3.1.0 in /usr/local/lib/python3.12/dist-packages (from scikit-learn) (3.6.0)
Requirement already satisfied: absl-py>=1.0.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (1.4.0)
Requirement already satisfied: astunparse>=1.6.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (1.6.3)
Requirement already satisfied: flatbuffers>=24.3.25 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (25.12.19)
Requirement already satisfied: gast!=0.5.0,!=0.5.1,!=0.5.2,>=0.2.1 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (0.7.0)
Requirement already satisfied: google_pasta>=0.1.1 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (0.2.0)
Requirement already satisfied: libclang>=13.0.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (18.1.1)
Requirement already satisfied: opt_einsum>=2.3.2 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (3.4.0)
Requirement already satisfied: protobuf>=5.28.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (5.29.6)
Requirement already satisfied: requests<3,>=2.21.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (2.32.4)
Requirement already satisfied: setuptools in /usr/local/lib/python3.12/dist-packages (from tensorflow) (75.2.0)
Requirement already satisfied: six>=1.12.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (1.17.0)
Requirement already satisfied: termcolor>=1.1.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (3.3.0)
Requirement already satisfied: typing_extensions>=3.6.6 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (4.15.0)
Requirement already satisfied: wrapt>=1.11.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (2.1.2)
Requirement already satisfied: grpcio<2.0,>=1.24.3 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (1.80.0)
Requirement already satisfied: tensorboard~=2.20.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (2.20.0)
Requirement already satisfied: keras>=3.10.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (3.13.2)
Requirement already satisfied: h5py>=3.11.0 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (3.16.0)
Requirement already satisfied: ml_dtypes<1.0.0,>=0.5.1 in /usr/local/lib/python3.12/dist-packages (from tensorflow) (0.5.4)
Requirement already satisfied: wheel<1.0,>=0.23.0 in /usr/local/lib/python3.12/dist-packages (from astunparse>=1.6.0->tensorflow) (0.47.0)
Requirement already satisfied: rich in /usr/local/lib/python3.12/dist-packages (from keras>=3.10.0->tensorflow) (13.9.4)
Requirement already satisfied: namex in /usr/local/lib/python3.12/dist-packages (from keras>=3.10.0->tensorflow) (0.1.0)
Requirement already satisfied: optree in /usr/local/lib/python3.12/dist-packages (from keras>=3.10.0->tensorflow) (0.19.0)
Requirement already satisfied: charset_normalizer<4,>=2 in /usr/local/lib/python3.12/dist-packages (from requests<3,>=2.21.0->tensorflow) (3.4.7)
Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.12/dist-packages (from requests<3,>=2.21.0->tensorflow) (3.13)
Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.12/dist-packages (from requests<3,>=2.21.0->tensorflow) (2.5.0)
Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.12/dist-packages (from requests<3,>=2.21.0->tensorflow) (2026.4.22)
Requirement already satisfied: markdown>=2.6.8 in /usr/local/lib/python3.12/dist-packages (from tensorboard~=2.20.0->tensorflow) (3.10.2)
Requirement already satisfied: pillow in /usr/local/lib/python3.12/dist-packages (from tensorboard~=2.20.0->tensorflow) (11.3.0)
Requirement already satisfied: tensorboard-data-server<0.8.0,>=0.7.0 in /usr/local/lib/python3.12/dist-packages (from tensorboard~=2.20.0->tensorflow) (0.7.2)
Requirement already satisfied: werkzeug>=1.0.1 in /usr/local/lib/python3.12/dist-packages (from tensorboard~=2.20.0->tensorflow) (3.1.8)
Requirement already satisfied: markupsafe>=2.1.1 in /usr/local/lib/python3.12/dist-packages (from werkzeug>=1.0.1->tensorboard~=2.20.0->tensorflow) (3.0.3)
Requirement already satisfied: markdown-it-py>=2.2.0 in /usr/local/lib/python3.12/dist-packages (from rich->keras>=3.10.0->tensorflow) (4.0.0)
Requirement already satisfied: pygments<3.0.0,>=2.13.0 in /usr/local/lib/python3.12/dist-packages (from rich->keras>=3.10.0->tensorflow) (2.20.0)
Requirement already satisfied: mdurl~=0.1 in /usr/local/lib/python3.12/dist-packages (from markdown-it-py>=2.2.0->rich->keras>=3.10.0->tensorflow) (0.1.2)

2. Library Imports

[ ]
import warnings; warnings.filterwarnings("ignore")
import pandas as pd
import numpy as np
import plotly.graph_objects as go
from plotly.subplots import make_subplots
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error, mean_absolute_error
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

3. Strategy Overview

Long Short-Term Memory (LSTM) networks are a class of recurrent neural networks (RNN) designed to capture temporal dependencies in sequential data. They are well-suited for price time-series because they maintain a cell state that can retain information over hundreds of timesteps.

Architecture:

  • Input layer: sliding window of seq_length bars × feature count.
  • LSTM layers with dropout regularisation to prevent overfitting.
  • Dense output layer: single neuron predicting the next bar's normalised close price.

Training procedure:

  1. Normalise close prices to [0, 1] using MinMaxScaler.
  2. Construct overlapping sequences of length seq_length.
  3. Train with early stopping on validation loss.
  4. Inverse-transform predictions to original price scale.

Signal derivation: Predicted close > current close → Buy (+1); < current close → Sell (−1).

Limitation: LSTMs require GPU resources for large datasets and hyperparameter tuning; on CPU they are slow. Random-walk synthetic data has minimal autocorrelation, so validation loss will be near-random.

4. Data Generation

[ ]
def generate_data(periods: int) -> pd.DataFrame:
    """
    Generate synthetic OHLCV price data using a geometric random walk.

    Parameters
    ----------
    periods : int
        Number of 1-minute bars to generate.

    Returns
    -------
    pd.DataFrame
        DataFrame with columns: open, high, low, close, volume, datetime.
    """
    start_date     = pd.to_datetime("2024-01-01 00:00:00+00:00")
    datetime_index = pd.date_range(start_date, periods=periods, freq="1min", tz="UTC")
    price_data = []
    last_close = 42000
    for i in range(periods):
        open_price  = last_close + np.random.normal(0, last_close * 0.0005)
        close_price = open_price + np.random.normal(0, last_close * 0.005)
        body_high   = max(open_price, close_price)
        body_low    = min(open_price, close_price)
        high_price  = max(body_high + abs(np.random.normal(0, last_close * 0.002)), open_price, close_price)
        low_price   = min(body_low  - abs(np.random.normal(0, last_close * 0.002)), open_price, close_price)
        if high_price < low_price:
            high_price, low_price = low_price, high_price
        price_data.append({
            "open":  max(1, int(open_price)),
            "high":  max(1, int(high_price)),
            "low":   max(1, int(low_price)),
            "close": max(1, int(close_price)),
        })
        last_close = close_price
    df = pd.DataFrame(price_data, index=datetime_index)
    df.index.name = "datetime"
    df["volume"]   = np.random.uniform(100.0, 500.0, periods)
    df["datetime"] = df.index.to_series()
    return df.reset_index(drop=True)

df = generate_data(500)
display(df.head())
open high low close volume datetime
0 42016 42052 41727 41781 107.627914 2024-01-01 00:00:00+00:00
1 41767 41866 41675 41691 410.389430 2024-01-01 00:01:00+00:00
2 41701 41706 41269 41349 428.486384 2024-01-01 00:02:00+00:00
3 41340 41350 41182 41185 444.410807 2024-01-01 00:03:00+00:00
4 41190 41467 41075 41440 262.143437 2024-01-01 00:04:00+00:00

5. LSTM Model

[ ]
def lstm_model(
    df: pd.DataFrame,
    seq_length: int = 30,
    epochs: int = 50,
    batch_size: int = 32,
    test_size: float = 0.2,
) -> tuple:
    """
    Build, train, and evaluate an LSTM model for next-bar close price prediction.

    Core logic
    ----------
    1. Extract and normalise the close price series with MinMaxScaler.
    2. Construct (X, y) pairs: X = sliding window of seq_length bars,
       y = the close price of the bar immediately following the window.
    3. Split chronologically into train/test sets.
    4. Define a two-layer LSTM with dropout, compiled with Adam and MSE loss.
    5. Train with early stopping (monitor val_loss, patience=10).
    6. Inverse-transform predictions and compute error metrics.

    Parameters
    ----------
    df : pd.DataFrame   OHLCV DataFrame with 'close' column.
    seq_length : int    Number of historical bars per input sequence.
    epochs : int        Maximum training epochs.
    batch_size : int    Mini-batch size.
    test_size : float   Fraction of data reserved for testing.

    Returns
    -------
    tuple
        (model, predictions, y_test_inv, scaler, history)
    """
    df = df.copy().sort_values("datetime", ignore_index=True)
    close = df["close"].values.reshape(-1, 1)

    # ── Normalisation ─────────────────────────────────────────────────────────
    scaler = MinMaxScaler(feature_range=(0, 1))
    scaled = scaler.fit_transform(close)

    # ── Sequence construction ─────────────────────────────────────────────────
    X, y = [], []
    for i in range(seq_length, len(scaled)):
        X.append(scaled[i - seq_length: i, 0])   # lookback window
        y.append(scaled[i, 0])                    # next bar target
    X, y = np.array(X), np.array(y)
    X = X.reshape(X.shape[0], X.shape[1], 1)      # (samples, timesteps, features)

    # ── Train/test split ──────────────────────────────────────────────────────
    split   = int(len(X) * (1 - test_size))
    X_train, X_test = X[:split], X[split:]
    y_train, y_test = y[:split], y[split:]

    # ── Model architecture ────────────────────────────────────────────────────
    model = Sequential([
        LSTM(64, return_sequences=True, input_shape=(seq_length, 1)),
        Dropout(0.2),
        LSTM(32, return_sequences=False),
        Dropout(0.2),
        Dense(1),
    ])
    model.compile(optimizer="adam", loss="mse")
    model.summary()

    # ── Training ──────────────────────────────────────────────────────────────
    es = EarlyStopping(monitor="val_loss", patience=10, restore_best_weights=True)
    history = model.fit(
        X_train, y_train,
        validation_split=0.1,
        epochs=epochs,
        batch_size=batch_size,
        callbacks=[es],
        verbose=1,
    )

    # ── Inference and inverse transform ──────────────────────────────────────
    preds    = model.predict(X_test)
    preds_inv = scaler.inverse_transform(preds)
    y_inv     = scaler.inverse_transform(y_test.reshape(-1, 1))

    rmse = np.sqrt(mean_squared_error(y_inv, preds_inv))
    mae  = mean_absolute_error(y_inv, preds_inv)
    print(f"\nTest RMSE: {rmse:.2f}  |  MAE: {mae:.2f}")

    return model, preds_inv, y_inv, scaler, history

model, preds, actuals, scaler, history = lstm_model(df, seq_length=30, epochs=30)
Model: "sequential"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ lstm (LSTM)                     │ (None, 30, 64)         │        16,896 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 30, 64)         │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ lstm_1 (LSTM)                   │ (None, 32)             │        12,416 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout_1 (Dropout)             │ (None, 32)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense (Dense)                   │ (None, 1)              │            33 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 29,345 (114.63 KB)

 Trainable params: 29,345 (114.63 KB)

 Non-trainable params: 0 (0.00 B)

Epoch 1/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 9s 235ms/step - loss: 0.0330 - val_loss: 0.0717
Epoch 2/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 2s 140ms/step - loss: 0.0103 - val_loss: 0.1011
Epoch 3/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 102ms/step - loss: 0.0071 - val_loss: 0.0639
Epoch 4/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 84ms/step - loss: 0.0066 - val_loss: 0.0774
Epoch 5/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 104ms/step - loss: 0.0066 - val_loss: 0.0611
Epoch 6/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 33ms/step - loss: 0.0067 - val_loss: 0.0528
Epoch 7/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 36ms/step - loss: 0.0058 - val_loss: 0.0492
Epoch 8/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 34ms/step - loss: 0.0061 - val_loss: 0.0388
Epoch 9/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0062 - val_loss: 0.0537
Epoch 10/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 35ms/step - loss: 0.0057 - val_loss: 0.0346
Epoch 11/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0052 - val_loss: 0.0291
Epoch 12/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 35ms/step - loss: 0.0059 - val_loss: 0.0341
Epoch 13/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 33ms/step - loss: 0.0049 - val_loss: 0.0363
Epoch 14/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 38ms/step - loss: 0.0052 - val_loss: 0.0226
Epoch 15/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 34ms/step - loss: 0.0053 - val_loss: 0.0214
Epoch 16/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0056 - val_loss: 0.0194
Epoch 17/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 36ms/step - loss: 0.0052 - val_loss: 0.0433
Epoch 18/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0056 - val_loss: 0.0366
Epoch 19/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0051 - val_loss: 0.0167
Epoch 20/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0042 - val_loss: 0.0198
Epoch 21/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 32ms/step - loss: 0.0045 - val_loss: 0.0314
Epoch 22/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 35ms/step - loss: 0.0045 - val_loss: 0.0203
Epoch 23/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 55ms/step - loss: 0.0043 - val_loss: 0.0129
Epoch 24/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 57ms/step - loss: 0.0041 - val_loss: 0.0166
Epoch 25/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 1s 57ms/step - loss: 0.0045 - val_loss: 0.0175
Epoch 26/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 34ms/step - loss: 0.0044 - val_loss: 0.0164
Epoch 27/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 35ms/step - loss: 0.0044 - val_loss: 0.0157
Epoch 28/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0040 - val_loss: 0.0087
Epoch 29/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 32ms/step - loss: 0.0041 - val_loss: 0.0121
Epoch 30/30
11/11 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step - loss: 0.0037 - val_loss: 0.0116
3/3 ━━━━━━━━━━━━━━━━━━━━ 1s 152ms/step

Test RMSE: 453.48  |  MAE: 365.23

6. Visualization — Training Loss and Predictions

[ ]
fig = make_subplots(rows=1, cols=2,
    subplot_titles=["Training / Validation Loss", "Predicted vs Actual Close"])

fig.add_trace(go.Scatter(y=history.history["loss"],     name="Train Loss",
    line=dict(color="blue")),  row=1, col=1)
fig.add_trace(go.Scatter(y=history.history["val_loss"], name="Val Loss",
    line=dict(color="orange")), row=1, col=1)

fig.add_trace(go.Scatter(y=actuals[:, 0], name="Actual",    line=dict(color="blue")),  row=1, col=2)
fig.add_trace(go.Scatter(y=preds[:, 0],   name="Predicted", line=dict(color="red",  dash="dash")), row=1, col=2)

fig.update_layout(title_text="LSTM Price Prediction", height=500)
fig.show()

Conclusion

This notebook demonstrates how to build and train an LSTM model for price prediction using synthetic data. Key steps include data generation, normalization, sequence construction, model definition, training with early stopping, and visualization of results. While LSTMs are powerful for time-series forecasting, their effectiveness depends on data characteristics and proper tuning.

LSTM Model · BitPredict