Harwinder Sekhon
01baef48ff
TF-IDF baseline (news_sentiment.py, models/xauusd_sentiment.joblib) and the two-head DeBERTa-v3-small transformer (transformer/, news_transformer.py) with its Google News crawler and DeepSeek labeller (data_collect/), the golden-set and archive evals (eval/), and the 2026-09-22 and 2026-09-27 results and progress notes. Transformer weights (model.pt, model.onnx) and the crawl database are not tracked; rebuild with transformer/train.py and transformer/export_onnx.py. The DeepSeek label cache in data/ is tracked.
274 lines
5.1 KiB
JSON
274 lines
5.1 KiB
JSON
{
|
|
"band": 0.15,
|
|
"band_sweep": {
|
|
"0.05": {
|
|
"hits": 51,
|
|
"n": 78,
|
|
"neutral_predicted": 9
|
|
},
|
|
"0.1": {
|
|
"hits": 52,
|
|
"n": 78,
|
|
"neutral_predicted": 13
|
|
},
|
|
"0.15": {
|
|
"hits": 50,
|
|
"n": 78,
|
|
"neutral_predicted": 20
|
|
},
|
|
"0.2": {
|
|
"hits": 45,
|
|
"n": 78,
|
|
"neutral_predicted": 26
|
|
},
|
|
"0.3": {
|
|
"hits": 38,
|
|
"n": 78,
|
|
"neutral_predicted": 40
|
|
},
|
|
"0.5": {
|
|
"hits": 28,
|
|
"n": 78,
|
|
"neutral_predicted": 53
|
|
}
|
|
},
|
|
"blocks": {
|
|
"DeepSeek live (xau through band \u00b10.15)": {
|
|
"accuracy": 1.0,
|
|
"confusion": [
|
|
[
|
|
32,
|
|
0,
|
|
0
|
|
],
|
|
[
|
|
0,
|
|
8,
|
|
0
|
|
],
|
|
[
|
|
0,
|
|
0,
|
|
38
|
|
]
|
|
],
|
|
"directional_accuracy": 1.0,
|
|
"directional_hits": 70,
|
|
"directional_n": 70,
|
|
"hits": 78,
|
|
"macro_f1": 1.0,
|
|
"n": 78,
|
|
"sign_agreement": 1.0,
|
|
"sign_hits": 70,
|
|
"sign_n": 70
|
|
},
|
|
"keyword fallback (band \u00b10.15)": {
|
|
"accuracy": 0.3076923076923077,
|
|
"confusion": [
|
|
[
|
|
6,
|
|
15,
|
|
11
|
|
],
|
|
[
|
|
0,
|
|
8,
|
|
0
|
|
],
|
|
[
|
|
12,
|
|
16,
|
|
10
|
|
]
|
|
],
|
|
"directional_accuracy": 0.22857142857142856,
|
|
"directional_hits": 16,
|
|
"directional_n": 70,
|
|
"hits": 24,
|
|
"macro_f1": 0.30646952758745044,
|
|
"n": 78,
|
|
"sign_agreement": 0.41025641025641024,
|
|
"sign_hits": 16,
|
|
"sign_n": 39
|
|
},
|
|
"model (native argmax label)": {
|
|
"accuracy": 0.6153846153846154,
|
|
"confusion": [
|
|
[
|
|
16,
|
|
5,
|
|
11
|
|
],
|
|
[
|
|
0,
|
|
8,
|
|
0
|
|
],
|
|
[
|
|
6,
|
|
8,
|
|
24
|
|
]
|
|
],
|
|
"directional_accuracy": 0.5714285714285714,
|
|
"directional_hits": 40,
|
|
"directional_n": 70,
|
|
"hits": 48,
|
|
"macro_f1": 0.6006169923663232,
|
|
"n": 78,
|
|
"sign_agreement": 0.7017543859649122,
|
|
"sign_hits": 40,
|
|
"sign_n": 57
|
|
},
|
|
"model (score through band \u00b10.15)": {
|
|
"accuracy": 0.6410256410256411,
|
|
"confusion": [
|
|
[
|
|
16,
|
|
5,
|
|
11
|
|
],
|
|
[
|
|
0,
|
|
8,
|
|
0
|
|
],
|
|
[
|
|
5,
|
|
7,
|
|
26
|
|
]
|
|
],
|
|
"directional_accuracy": 0.6,
|
|
"directional_hits": 42,
|
|
"directional_n": 70,
|
|
"hits": 50,
|
|
"macro_f1": 0.6228451632225217,
|
|
"n": 78,
|
|
"sign_agreement": 0.7241379310344828,
|
|
"sign_hits": 42,
|
|
"sign_n": 58
|
|
}
|
|
},
|
|
"deepseek": {
|
|
"batch_ms_max": null,
|
|
"batch_ms_mean": null,
|
|
"cache_hits": 78,
|
|
"calls": 0,
|
|
"failed_calls": 0,
|
|
"headlines_sent": 0,
|
|
"model": "deepseek-v4-flash",
|
|
"per_headline_ms": null,
|
|
"prompt_sha1": "d32e04c6a87d019f326f111b39c99b6dd8ce794d"
|
|
},
|
|
"deepseek_scored": 78,
|
|
"generated_at": "20260927_1859_transformer",
|
|
"mcnemar": {
|
|
"model_vs_deepseek": {
|
|
"deepseek_only_right": 30,
|
|
"model_only_right": 0,
|
|
"p": 1.862645149230957e-09
|
|
},
|
|
"model_vs_keyword": {
|
|
"keyword_only_right": 6,
|
|
"model_only_right": 30,
|
|
"p": 6.960169412195683e-05
|
|
}
|
|
},
|
|
"model": {
|
|
"final": false,
|
|
"model_path": "/home/harwinder/Development/ML-NEWS/models/xau_transformer",
|
|
"model_type": "two_head_transformer:microsoft/deberta-v3-small",
|
|
"oos_from": "2026-09-18",
|
|
"relevance_threshold": 0.2,
|
|
"trained_at": "2026-09-27T18:58:27.596998+00:00",
|
|
"training_examples": 2402
|
|
},
|
|
"model_latency_ms": {
|
|
"max": 10.71,
|
|
"mean": 8.942564102564102,
|
|
"median": 8.87
|
|
},
|
|
"n": 78,
|
|
"per_category": {
|
|
"distractors": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 8,
|
|
"model": 8,
|
|
"n": 8
|
|
},
|
|
"dovish_fed": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 7,
|
|
"model": 4,
|
|
"n": 8
|
|
},
|
|
"gold_direct": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 4,
|
|
"model": 6,
|
|
"n": 8
|
|
},
|
|
"hawkish_fed": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 2,
|
|
"model": 7,
|
|
"n": 8
|
|
},
|
|
"inverted_metrics": {
|
|
"deepseek": 6,
|
|
"deepseek_n": 6,
|
|
"keyword": 0,
|
|
"model": 1,
|
|
"n": 6
|
|
},
|
|
"negation_traps": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 2,
|
|
"model": 3,
|
|
"n": 8
|
|
},
|
|
"real_yields": {
|
|
"deepseek": 6,
|
|
"deepseek_n": 6,
|
|
"keyword": 1,
|
|
"model": 5,
|
|
"n": 6
|
|
},
|
|
"risk_on": {
|
|
"deepseek": 4,
|
|
"deepseek_n": 4,
|
|
"keyword": 0,
|
|
"model": 0,
|
|
"n": 4
|
|
},
|
|
"safe_haven": {
|
|
"deepseek": 8,
|
|
"deepseek_n": 8,
|
|
"keyword": 0,
|
|
"model": 8,
|
|
"n": 8
|
|
},
|
|
"strong_usd": {
|
|
"deepseek": 7,
|
|
"deepseek_n": 7,
|
|
"keyword": 0,
|
|
"model": 1,
|
|
"n": 7
|
|
},
|
|
"weak_usd": {
|
|
"deepseek": 7,
|
|
"deepseek_n": 7,
|
|
"keyword": 0,
|
|
"model": 5,
|
|
"n": 7
|
|
}
|
|
},
|
|
"spearman_model_vs_deepseek": 0.5012357134215089
|
|
}
|