Neural Asset Forecaster: A Deep Learning-Driven Multi-Asset Price Prediction System

This article introduces Neural Asset Forecaster, an automated price prediction system based on an ensemble of multiple advanced neural network models. The system provides 30-day rolling forecasts for Bitcoin, Ethereum, gold, and silver with daily automatic updates. Its core highlight is the adoption of an ensemble strategy using five complementary models such as TSMixer and NBEATS, implemented via the NeuralForecast framework, aiming to provide data-driven references and a reusable framework for investors, researchers, and developers.

In financial markets, accurately predicting asset price trends is a core goal. Traditional methods like ARIMA have limitations when dealing with non-linear, highly volatile financial data. Deep learning technology has made significant breakthroughs in time series prediction, especially suitable for the high-dimensional features and complex patterns of financial data. The Neural Asset Forecaster project integrates advanced neural network models to build a fully automated multi-asset prediction system, demonstrating the application of deep learning in the financial field and providing a reusable and extensible prediction framework.

The project adopts a modular design centered around the NeuralForecast library (a deep learning framework specifically for time series prediction). The system runs automatically every midnight UTC and generates 30-day future forecasts. Supported assets include: Gold (GC=F), Silver (SI=F), Bitcoin (BTC-USD), Ethereum (ETH-USD), covering traditional precious metals and cryptocurrencies. It can capture the correlations and differences between different markets, providing multi-dimensional references for portfolio management.

The project uses an ensemble of five complementary neural network models for prediction:

1. TSMixer: A lightweight MLP-based model that captures dependencies through channel mixing and temporal mixing. It is configured with 4 residual blocks, a 128-dimensional feedforward layer, a dropout rate of 0.5, an input window of 504 days, and uses RevIN to handle distribution shifts.
2. NBEATS: A pure fully connected architecture with stacked trend and seasonal stacks (3 blocks each, 512-dimensional MLP units), explicitly decomposing trend and seasonal components.
3. NHITS: An improved version of NBEATS that introduces hierarchical interpolation and frequency downsampling ([24,12,1]) to achieve multi-scale modeling.
4. MLP: A baseline model with 2 hidden layers (512 neurons each), an input window of 252 days, used as a comparison baseline.
5. TiDE: An encoder-decoder architecture with dense connections and temporal attention. It is configured with a 512-dimensional hidden layer, a 32-dimensional decoder output, a 128-dimensional temporal decoder, and a dropout rate of 0.5.

The system's data flow is concise and efficient:

1. Data Collection: Historical price data from January 1, 2018 onwards is obtained from Yahoo Finance via yfinance.
2. Preprocessing: Data is standardized using a robust scaler and formatted into a three-column structure: ds (date), y (target value), unique_id (sequence identifier).
3. Training and Prediction: Five models are trained separately for each asset, using 30 days as the validation set size, and generate 30-day future predictions.
4. Post-processing: Prediction results are Gaussian-smoothed (sigma=2) to reduce short-term noise.
5. Visualization: Comparative charts of price trends and prediction intervals are automatically generated and saved in PNG format. The entire process is automatically run daily via GitHub Actions without manual intervention.

The application value of Neural Asset Forecaster includes:

• Investors: Provides data-driven price trend references to assist in trading decisions and risk management (as a supplement to technical analysis and fundamental analysis).
• Researchers: Demonstrates the application of modern deep learning models in financial time series prediction and provides reproducible benchmark implementations.
• Developers: The code has a clear structure and high modularity, making it easy to extend new asset classes or try other prediction models. It is an excellent example for learning time series prediction.

Any price prediction system has inherent limitations:

• Market Unpredictability: Financial markets are influenced by complex factors such as macroeconomics, policy changes, and unexpected events, which are difficult to fully incorporate into historical price data modeling.
• Model Overfitting Risk: Deep learning models have large capacity and may overfit historical patterns, leading to reduced generalization ability for future data.
• Black Box Nature: The decision-making process of neural network models lacks interpretability, making it difficult for users to understand the logic behind prediction results. It is recommended to use this system as an auxiliary tool and make comprehensive judgments in combination with other analysis methods, rather than blindly following prediction signals for trading.

Neural Asset Forecaster is a technologically advanced and well-designed automated prediction system that successfully applies multiple SOTA neural network models to long-term price prediction of financial assets. Through the ensemble strategy and automated operation and maintenance, it continuously produces high-quality forecasts, providing references for participants in the cryptocurrency and precious metal markets. Future expansion directions include: introducing more external features (trading volume, market sentiment indicators, macroeconomic data), trying advanced model architectures such as Transformer, and developing real-time prediction and early warning functions. This project has provided a solid starting point for the application of deep learning in the financial field.