Workers
Overview
Workers run machine learning models that generate predictions on network topics. The Allora Network uses consensus mechanisms to combine predictions from multiple workers, reward accurate models, and synthesize final inferences for consumers.
As a worker, you deploy models that submit inferences to specific topics (e.g., ETH price prediction, market volatility forecasting). Your rewards are determined by prediction accuracy relative to ground truth, measured through loss functions and weighted by consensus mechanisms.
How It Works
The network crowdsources financial predictions through machine learning models. When you run a worker node:
- Select topics matching your model's capabilities
- Submit inferences at regular intervals defined by the topic
- Network scores your predictions against ground truth data
- Earn rewards proportional to accuracy relative to other workers
The consensus mechanism combines multiple worker predictions using techniques like EMA (Exponential Moving Average) to weight recent performance more heavily, allowing the network to adapt quickly to model accuracy changes.
Worker Responsibilities
- Deploy and maintain inference models
- Submit predictions according to topic schedules
- Monitor model performance and accuracy
- Retrain models as needed to maintain competitive accuracy
Technical Requirements
Infrastructure
- Computing resources sufficient for model inference
- Reliable network connectivity to Allora RPC nodes
- Storage for model weights and training data
- Monitoring setup to track inference submission and rewards
Integration
Workers integrate via the worker node software, which handles:
- Network communication and authentication
- Inference submission timing
- Reward tracking and distribution
- Connection management to Allora blockchain
See system requirements for detailed specifications.
Model Development
Data and Training
Focus on high-quality training data relevant to your target topics. Financial prediction models benefit from:
- Clean, preprocessed price and market data
- Appropriate feature engineering for the prediction task
- Robust validation procedures to avoid overfitting
- Regular retraining as new data becomes available
Model Selection
Choose algorithms suited to your prediction task. Time-series forecasting for price prediction typically uses models like:
- LSTM and GRU networks for sequence prediction
- Transformer architectures for attention-based forecasting
- Ensemble methods combining multiple approaches
- Classical models (ARIMA, Prophet) as baselines
Performance Optimization
Continuously monitor your model's accuracy on the network. The consensus mechanism means you're competing against other workers, so consistent accuracy is crucial for maximizing rewards.
Analyze prediction errors to identify:
- Systematic biases in your model
- Market conditions where accuracy drops
- Opportunities for feature or architecture improvements
Deployment
Deploy using Docker for the fastest setup, or build from source for custom configurations.
After deployment, query worker data to monitor performance and view available topics to select prediction targets.