File size: 2,961 Bytes
4a4f066
6cbbcc4
 
 
 
 
 
4a4f066
6cbbcc4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
---
tags:
  - tabular-classification
  - sme
  - credit-risk
  - calibrated-classification
library_name: scikit-learn
---

# PACE-SME Proposed Model

This repository contains the fitted PACE-SME proposed two-component cross-fitted simplex classifier, its inference helper, feature schema, runtime dependencies, and aggregated evaluation assets.

## Intended use

Research-only retrospective classification against the `target_missed_payment` label in the supplied Bangladesh SME dataset. The current data provenance and prediction-time horizon are not fully verified, so this artifact must not be represented as a validated prospective early-warning system.

The model returns the positive-class score and a threshold-based label. It is a research screening signal only. A qualified human reviewer must assess the original records and context. Do not use this score as the sole basis for credit approval, denial, pricing, restructuring, or another consequential borrower decision. It does not estimate causal restructuring effects or identify an optimal intervention.

## Model details

- Architecture: two-component cross-fitted probability simplex.
- Components selected using development-only grouped out-of-fold results: CatBoost_Unweighted, LightGBM_Unweighted.
- Calibration: isotonic.
- Validation-selected decision threshold: 0.22682374.
- Input features: see `feature_schema.json` and `model_config.json`.
- Training and locked-test rows: 960 and 240.
- Current nested-CV status: COMPLETED.

## Inference

Install `requirements.txt`, then load the model with:

```python
import json
from pace_inference import load_predictor
predictor = load_predictor('.')
with open('example_input.json', encoding='utf-8') as input_file:
    record = json.load(input_file)
result = predictor.predict(record)
print(result)
```

Every feature in `feature_schema.json` is required. Extra input columns are ignored. Missing feature columns raise an error. Feature values must follow the schema and preprocessing learned during training. Replace each null placeholder in `example_input.json` with a valid value before prediction.

## Evaluation assets

Aggregated validation and locked-test metrics are in `performance/metrics.json`; comparison tables are in the same directory; generated ROC, precision-recall, calibration, confusion-matrix, decision-curve, threshold-sensitivity, and capacity plots are in `figures/`. Metrics are produced by the current notebook run. They must be interpreted with their dataset, split, and evidence limitations.

## Reproducibility and limitations

The fitted artifact is serialized with cloudpickle and depends on the exact package versions recorded in `requirements.txt`. Do not load pickle-based model files from untrusted sources. Raw datasets, row-level records, and row-level test predictions are not included in this repository. Dataset license and redistribution permissions must be checked before making this repository public.