Abstract

This report details the end-to-end architectural pipeline used to develop and validate the Bitcoin (BTC) trading component of the BOT FERRARI system. The core challenge was designing a robust predictive model capable of achieving exceptional risk-adjusted returns in the highly volatile cryptocurrency market, specifically targeting the recent 2.3-year market micro-trend. The solution utilizes a hybrid 12-feature Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model, whose stability was rigorously confirmed through Walk-Forward Optimization (WFO) and Hyperband tuning. The final candidate, the MLM-12 model, achieved an optimal Average Out-of-Sample Sharpe Ratio of 5.19 and a total compounded return of 636.27%, proving its superior efficacy and suitability for automated deployment.

Introduction

Historically, financial market prediction suffered from relying on linear models that failed to capture the non-stationary, non-linear dynamics of high-frequency cryptocurrency data. The intelligence of the BTC Bot bypasses these limitations by employing deep learning. Furthermore, in trading, the significant price movements that generate profitable signals (Buy/Sell) are statistically rare compared to periods of low activity (Hold). Addressing this fundamental class imbalance (where Hold signals dominate) is critical for ensuring the model does not become biased toward the passive 'Hold' signal.

1. Data Foundation: The 12-Year Historical Core

The foundation of any robust algorithmic trading strategy is clean, comprehensive data. To mitigate overfitting and ensure the model learns macro-cyclical patterns, the initial step involved curating a 12-year archive of Bitcoin's hourly OHLCV (Open, High, Low, Close, Volume) data, meticulously stored in an SQLite database.

The pipeline utilizes two distinct tables within the ohlcv_data_BTC.db SQLite file to enforce data scope segregation:

• btcusd_1h_data_12y: This table contains the entire 12-year history and is used exclusively to train the initial MLM-12F deep learning model, providing it with depth of knowledge across multiple market cycles.


• btcusd_1h_data: This table contains the recent 2.3 years of data and forms the basis for the WFO-Tuner-Framework, ensuring that parameter optimization and final validation are strictly tethered to current market microstructure and volatility.

Crucially, while the model training leverages the whole 12-year history (btcusd_1h_data_12y) to capture broad market regimes, the critical validation and optimization stages focus solely on the most recent 2.3 years (btcusd_1h_data) of data. This methodology—training on macro history but validating against current trends—ensures the model's knowledge is deep while its trading parameters remain relevant to current market microstructure.

2. The Predictive Core: Hybrid CNN-LSTM Architecture (12 Features)

The prediction system uses a specialized deep learning architecture designed for time series analysis:

• Hybrid Structure: A combination of CNN layers (to identify local price patterns and spatial features over the look-back window) and stacked LSTM layers (to capture long-term temporal dependencies).


• Feature Simplicity (The Winning Strategy): The final, best-performing model utilizes a reduced set of 12 input features (MLM-12). This feature set includes the base OHLCV data alongside key technical indicators (TIs) like RSI, MACD, BBands (excluding the Middle band), ATR, and OBV, but strategically omits more complex or highly correlated indicators (like specific EMAs).


• Data Preparation: The raw sequential data is segmented into training samples, each representing a 72-hour (3-day) look-back window. A hybrid resampling technique (RUS followed by SMOTE) is applied exclusively to the training set to address the significant class imbalance (Buy/Sell/Hold) inherent in financial data, enabling the model to effectively learn the minority 'Buy' and 'Sell' signals.

The model training process itself utilized callbacks, such as Early Stopping and ReduceLROnPlateau, applied to the validation loss to halt training when marginal improvements ceased, thereby proactively preventing model overfitting.

3. Rigorous Validation: Walk-Forward Optimization (WFO)

To transcend the critical flaw of backtest overfitting, the trading logic was subjected to Walk-Forward Optimization, combining the highest standards of financial rigour with advanced machine learning techniques:

• WFO Engine: The strategy was tested across 81 sequential folds within the target 2.3-year market scope. Each fold involved re-optimizing the trading parameters on an "in-sample" window and validating the results on the subsequent, unseen "out-of-sample" data.


• Hyperparameter Tuning: The Hyperband tuner was employed within the WFO loop. Its objective was explicitly set to maximize the Average Out-of-Sample Sharpe Ratio, thus ensuring that the final parameters prioritize risk-adjusted consistency over raw profit.


• Parameter Selection: The tuner optimized the entire risk management framework, including the prediction confidence threshold, volatility-based position sizing (leveraging ATR), and adaptive stop-loss multipliers.

4. Conclusion and Final Deployment Mandate

The completion of this rigorous Walk-Forward Validation delivers a decisive victory over the non-stationary nature of the cryptocurrency market. The stability check confirms the model operates at a standard of excellence rarely seen in volatile markets.

The MLM-12 is the final, proven candidate for the BTC-Bot component of BOT FERRARI. The validated 5.19 Sharpe Ratio signifies an extraordinary level of risk management and return generation, transforming the complex quantitative model into a robust, disciplined operational engine. The superior consistency of the 12-feature model guarantees the system is ready for immediate, high-conviction deployment as the flagship component of BOT FERRARI.

By FRANK MORALES

Keywords: Cryptocurrency, Open Source, Predictive Analytics

Follow Us On