Implementation:Shiyu coder Kronos Prediction Result Output

Knowledge Sources	Shiyu_coder_Kronos
Domains	Data_Persistence, Financial_Prediction, Serialization
Last Updated	2026-02-09 14:00 GMT

Overview

JSON output format specification for persisted prediction results generated by the Kronos WebUI, containing forecasted OHLCV candlestick data alongside ground truth for comparison analysis.

Description

Each time the WebUI executes a prediction via the /api/predict endpoint, the results are saved as a timestamped JSON file in webui/prediction_results/. The format captures the complete prediction context: input parameters, data summary, predicted candlesticks, actual (ground truth) candlesticks, and continuity analysis between predicted and actual values.

This is a Pattern Doc — it documents a data format rather than a callable API. The files are produced by save_prediction_results() in webui/app.py.

Usage

Reference this format when parsing or analyzing saved prediction outputs from the Kronos WebUI. Each JSON file is self-contained and includes all metadata needed to reproduce or evaluate the prediction.

Code Reference

Source Location

• Repository: Shiyu_coder_Kronos
• File: webui/app.py (save_prediction_results)
• Output Directory: webui/prediction_results/

Signature

{
  "timestamp": "2025-08-26T16:38:00.302387",
  "file_path": "/path/to/source_data.feather",
  "prediction_type": "Kronos model prediction (...)",
  "prediction_params": {
    "lookback": 400,
    "pred_len": 120,
    "temperature": 1.0,
    "top_p": 0.9,
    "sample_count": 1,
    "start_date": "2025-08-02T13:24"
  },
  "input_data_summary": {
    "rows": 400,
    "columns": ["open", "high", "low", "close", "volume"],
    "price_range": {
      "open": {"min": 112046.9, "max": 114231.95},
      "high": {"min": 112200.0, "max": 114260.31},
      "low": {"min": 111920.0, "max": 114085.61},
      "close": {"min": 112046.9, "max": 114231.95}
    },
    "last_values": {
      "open": 113769.56, "high": 113852.6,
      "low": 113731.29, "close": 113818.87
    }
  },
  "prediction_results": [
    {"timestamp": "...", "open": ..., "high": ..., "low": ..., "close": ..., "volume": ..., "amount": ...}
  ],
  "actual_data": [
    {"timestamp": "...", "open": ..., "high": ..., "low": ..., "close": ..., "volume": ..., "amount": ...}
  ],
  "analysis": {
    "continuity": {
      "last_prediction": {"open": ..., "high": ..., "low": ..., "close": ...},
      "first_actual": {"open": ..., "high": ..., "low": ..., "close": ...},
      "gaps": {"open_gap": ..., "high_gap": ..., "low_gap": ..., "close_gap": ...},
      "gap_percentages": {"open_gap_pct": ..., "high_gap_pct": ..., "low_gap_pct": ..., "close_gap_pct": ...}
    }
  }
}

Import

import json

# Load a prediction result file
with open("webui/prediction_results/prediction_20250826_163800.json") as f:
    result = json.load(f)

I/O Contract

Inputs

Name	Type	Required	Description
file_path	str	Yes	Source data file path used for prediction
prediction_type	str	Yes	Description string of prediction mode
prediction_results	list[dict]	Yes	List of predicted OHLCV candles with timestamps
actual_data	list[dict]	No	Ground truth candles for comparison (may be empty)
input_data	DataFrame	Yes	Historical input data (summarized in output)
prediction_params	dict	Yes	Parameters: lookback, pred_len, temperature, top_p, sample_count, start_date

Outputs

Name	Type	Description
JSON file	File	Timestamped file at webui/prediction_results/prediction_YYYYMMDD_HHMMSS.json
prediction_results	list[dict]	120 predicted OHLCV candles with ISO timestamps
actual_data	list[dict]	120 actual OHLCV candles (if ground truth available)
analysis.continuity	dict	Gap analysis between first predicted and first actual candle

Usage Examples

Loading and Analyzing a Prediction Result

import json
import pandas as pd

# Load saved prediction
with open("webui/prediction_results/prediction_20250826_163800.json") as f:
    result = json.load(f)

# Extract prediction parameters
params = result["prediction_params"]
print(f"Lookback: {params['lookback']}, Pred Length: {params['pred_len']}")
print(f"Temperature: {params['temperature']}, Top-p: {params['top_p']}")

# Convert predictions to DataFrame
pred_df = pd.DataFrame(result["prediction_results"])
pred_df["timestamp"] = pd.to_datetime(pred_df["timestamp"])
print(f"Predicted {len(pred_df)} candles")
print(f"Price range: {pred_df['close'].min():.2f} - {pred_df['close'].max():.2f}")

# Compare with actual data if available
if result["actual_data"]:
    actual_df = pd.DataFrame(result["actual_data"])
    actual_df["timestamp"] = pd.to_datetime(actual_df["timestamp"])

    # Check continuity gaps
    if "continuity" in result.get("analysis", {}):
        gaps = result["analysis"]["continuity"]["gap_percentages"]
        print(f"Close price gap: {gaps['close_gap_pct']:.2f}%")

Files Covered

This page documents the output format for all 29 prediction result files in webui/prediction_results/:

• prediction_20250826_163800.json through prediction_20250826_181932.json
• All files contain BTC/USDT 5-minute candlestick predictions with lookback=400 and pred_len=120

Related Pages

• Environment:Shiyu_coder_Kronos_PyTorch_CUDA_Environment