RavaPush

Key Concepts

Time Series Prediction

RavaPush analyzes data points arranged in chronological order (time series) and identifies patterns within them:

• Trends — Long-term directional changes
• Cyclicality — Repeating patterns
• Seasonality — Changes at fixed time intervals
• Anomalies — Data points deviating from normal patterns

Based on these patterns, the system generates probability distributions of future values and provides confidence intervals.

Zero-Shot Prediction

Due to pre-training on large-scale multi-domain data, RavaPush can be immediately applied to new time series:

• No need to train a model for each new dataset
• Can recognize common temporal patterns across domains
• Automatically improves as data accumulates

For specific applications, you can choose to fine-tune the model on your data to improve accuracy (typically 20-40% improvement).

Proactive Push

RavaPush doesn't just predict — it also proactively pushes alerts in the following situations:

• Deviation from historical patterns detected
• Predictions indicate future risks
• Preset thresholds reached
• Anomalous patterns match known risk signatures

The push system learns your response patterns to avoid excessive notifications.

Typical Applications

Tracking and Monitoring

Personal Habits

Resource Usage

Customer Behavior

Progress and Planning

Project Estimation

Capacity Planning

Inventory Management

Anomaly Detection

Quality Monitoring

Security Threats

Supply Chain Risk

Data Requirements

Accuracy and Limitations

Factors Affecting Accuracy

Data Volume and Quality

More data = better pattern recognition. Regular data is more predictable than irregular data.

Prediction Range

Short-term (1-7 days) is usually most accurate; long-term (30+ days) has significantly increased uncertainty.

Pattern Stability

Stable patterns are easy to predict; frequently changing patterns are difficult to predict.

Main Limitations

Cannot Predict Unprecedented Events

The system is based on historical patterns. If something has never happened before, or a fundamental change occurs, predictions will fail.

Correlation, Not Causation

The system finds statistical correlations but doesn't understand causal relationships. Changing underlying factors may invalidate predictions.

Requires Sufficient Historical Data

Sensitive to Data Quality

Missing values, inconsistent intervals, and measurement method changes all reduce accuracy.

Technical Details

Model Architecture

RavaPush uses a decoder-only Transformer architecture:

• Autoregressive Generation — Predicts the next value step by step, similar to language models
• Patch-based — Processes sequence in segments (32 time points)
• Causal Attention — Only looks at past information, preventing data leakage
• Maximum Context — 2048 time points

Pre-training

The model is pre-trained on over 100 billion real-world time series data points, covering multiple industries and application scenarios, including various time granularities (hourly to quarterly), identifying common patterns across domains. This enables the model to:

• Perform zero-shot prediction on new time series
• Be more robust to noise and outliers
• Quickly adapt to new scenarios

Confidence Intervals

Each prediction includes confidence intervals (10th to 90th percentile):

• Narrow interval — High confidence, clear patterns
• Wide interval — Low confidence, high uncertainty

Fine-tuning (Optional)

For scenarios with sufficient data (500+ points) and special patterns, you can fine-tune the model:

• Fine-tune on top of the foundation model
• Retain generalization ability while enhancing sensitivity to specific patterns
• Accuracy typically improves by 20-40%
• Fine-tuning time: hours to one day

Best Practices

Data Collection

• Maintain consistent measurement methods and time intervals
• Record reasons for anomalous events
• Avoid long data gaps
• Accumulate as much historical data as possible

Interpreting Predictions

• Focus on confidence intervals, not just point predictions
• Compare predictions with actual results to understand system reliability
• Understand that predictions are suggestions; the final decision is yours
• Regularly update data to improve predictions

When to Trust Predictions

Comparison with Other Methods

Role in RavaAI

RavaPush is one of three core components of RavaAI:

Synergy Example

Considering whether to accelerate project schedule:

FAQ

How much data is needed?

Minimum 30 points to start predicting, 100+ points for reliable results, 500+ points for fine-tuning.

How far ahead can it predict?

Technically any length, but accuracy decreases over time. Short-term (1-7 days) is most accurate; long-term (30+ days) has high uncertainty.

What types of data does it handle?

Numerical univariate time series with regular intervals. Does not support text, categorical data, or multivariate prediction.

How does it handle missing data?

Forward fill, interpolation, or exclusion. Missing data exceeding 10% reduces reliability.

Do predictions update automatically?

Yes. After new data is recorded, the model recalculates patterns and updates predictions.

How is push frequency controlled?

The system uses confidence thresholds, importance assessment, frequency control, and learning your response patterns to avoid excessive notifications. Adjustable in settings.

How long does fine-tuning take?

500-1000 points takes a few hours; 1000+ points may take half a day to one day.

Privacy

• Your data is used to generate personal predictions
• Fine-tuned models serve only you
• Not shared with other users
• Can be deleted at any time