Principle:Shiyu coder Kronos Prediction Result Persistence
| Knowledge Sources | |
|---|---|
| Domains | Data_Persistence, Financial_Prediction, Reproducibility |
| Last Updated | 2026-02-09 14:00 GMT |
Overview
Design pattern for persisting time-series prediction outputs as self-contained JSON files that capture the full prediction context for reproducibility and post-hoc analysis.
Description
Prediction Result Persistence addresses the need to store model prediction outputs in a structured, self-contained format. In financial time-series forecasting, predictions are ephemeral by nature — once generated, they must be compared against future actual data. This principle ensures that every prediction run creates a durable record containing:
- Input context — What data was fed to the model (summary statistics, not raw data)
- Generation parameters — Exact hyperparameters used (temperature, top_p, lookback window)
- Predicted values — The full sequence of forecasted candlesticks with timestamps
- Ground truth — Actual values from the same time period (if available at prediction time)
- Continuity analysis — Quantitative gap measurements between predicted and actual values
This enables retrospective evaluation of model performance across different market conditions and parameter settings without re-running inference.
Usage
Apply this principle when building prediction systems where users need to review, compare, or audit past predictions. It is essential when predictions are made ahead of time and need to be validated against subsequently observed data.
Theoretical Basis
The persistence pattern follows a Prediction Audit Trail design:
# Abstract pattern (NOT implementation code)
# Each prediction creates a complete audit record:
prediction_record = {
"when": timestamp_of_prediction,
"what": source_data_summary,
"how": generation_parameters,
"result": predicted_values,
"truth": actual_values_if_available,
"quality": gap_analysis_metrics
}
# File naming encodes time for natural ordering:
# prediction_YYYYMMDD_HHMMSS.json
save(prediction_record, unique_timestamped_path)
Key design properties:
- Self-contained — Each file includes all metadata needed to understand the prediction without external references
- Append-only — New predictions create new files; existing results are never modified
- Time-ordered — Filename timestamps enable chronological sorting and deduplication