Model Card: LPR-Oracle
1. Summary
LPR-Oracle is a deep learning-based time series classification model designed to predict adjustment events of China's Loan Prime Rate (LPR) solely through micro-market interest rate data. The model employs an ensemble learning strategy based on the InceptionTime architecture, moving away from traditional macro-fundamental factors to focus on mining non-linear patterns within the historical spread between Shibor and LPR. Evaluation results indicate that under F2-Score-oriented threshold optimization, the model demonstrates significant high-sensitivity characteristics, making it suitable as an Early Warning System (EWS) in interest rate risk management.
2. Model Description
2.1 Architecture Design
- Backbone Network: InceptionTime (Fawaz et al., 2020).
- Utilizes multi-scale 1D-CNN modules (Kernel sizes: 9, 19, 39) to extract short-, medium-, and long-term temporal dependencies in parallel.
- Introduces Bottleneck layers to reduce dimensionality and parameter count, preventing overfitting on small-sample datasets.
- Uses Global Average Pooling (GAP) instead of fully connected layers to enhance generalization capability.
- Ensemble Strategy:
- Deploys N=5 independently initialized models.
- Introduces seed diversity to reduce prediction variance.
- Employs a Soft-Voting mechanism in the output layer for probability averaging.
2.2 Input Space
The model relies exclusively on pure numerical high-frequency interbank market data, without incorporating any unstructured text or macroeconomic indicators:
- Feature Vector: Xt = [Spread_t, Momentum_t]
- Spread_t: The interest rate spread between 1-year LPR and 1-year Shibor.
- Momentum_t: The first-order difference sequence of Shibor.
- Lookback Window: T=60, capturing quarterly-level liquidity trends.
3. Training Methodology
- Loss Function: Binary Cross-Entropy (BCE) with Class Re-weighting. Penalty weights are applied to positive samples to address the sparsity of LPR change events (Class Imbalance).
- Threshold Calibration:
- Abandons the traditional 0.5 threshold.
- Adopts an F2-Score Maximization strategy: beta=2, meaning the weight of Recall is twice that of Precision.
- Theoretical Basis: In financial crisis contagion models, the marginal cost of Type II Error (missed risk) is significantly higher than that of Type I Error (false alarm).
4. Evaluation Results
Statistical evaluation based on the test set is as follows:
| Metric | Value | Statistical Interpretation |
|---|---|---|
| ROC-AUC | 0.6215 | The model demonstrates discriminative power superior to random guessing (0.5). |
| PR-AUC | 0.2006 | Reflects the extreme scarcity of positive samples (LPR changes). |
| Sensitivity | 0.9231 | Key metric. The model successfully identified 12 out of 13 change events, proving the effectiveness of Inception modules in capturing subtle market anomalies. |
| Precision | 0.1739 | Low precision is the direct cost of a high-recall strategy. The model tends to misclassify high-volatility market noise as signals. |
| Specificity | 0.2400 | The model has a weak ability to identify stable periods and shows a significant tendency to overreact. |
Confusion Matrix Analysis
- TP (12) vs FN (1): The extremely low miss rate indicates the model is highly reliable as a "filter."
- FP (57): The high false alarm rate suggests the model captures numerous market stress signals that do not ultimately translate into LPR adjustments. While statistically labeled as "errors," these signals may represent real liquidity tension in an economic sense.
5. Limitations & Discussion
- Information Asymmetry:
- Relying solely on Shibor/LPR values cannot capture exogenous variables such as "policy shocks" or "window guidance." The upper limit of ROC-AUC (0.62) is constrained by insufficient information entropy in the input data.
- Precision-Recall Trade-off:
- The current optimal threshold (0.9957) is in an extremely skewed state. This causes Accuracy (0.3409) to lose its reference value. This model should not serve as a direct trigger for automated trading systems but rather as a Decision Support System for analysts.
6. Conclusion
LPR-Oracle validates the feature extraction capabilities of deep convolutional neural networks under minimalist feature inputs. By sacrificing specificity to maximize sensitivity, the model establishes a high-sensitivity interest rate change monitoring mechanism suitable for risk-averse financial applications.