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.
26 lines
1.3 KiB
Plaintext
26 lines
1.3 KiB
Plaintext
DeepSeek deepseek-v4-flash, prompt d32e04c6; budget 20 calls
|
|
pools: precise 3137, broad 20533 unlabelled; picked 500
|
|
selected 500 unlabelled headlines
|
|
250 labelled, 10 calls (0 failed), 0 from cache
|
|
500 labelled, 20 calls (0 failed), 0 from cache
|
|
this run: 20 calls, 0 failed, 500 headlines labelled
|
|
1500 labelled headlines
|
|
year n bear irrel bull relevant mean xau
|
|
2009 4 0 4 0 0.0% +0.000
|
|
2010 6 0 6 0 0.0% +0.000
|
|
2011 9 0 9 0 0.0% +0.000
|
|
2012 2 0 2 0 0.0% +0.000
|
|
2015 8 0 8 0 0.0% +0.000
|
|
2016 5 0 5 0 0.0% +0.000
|
|
2017 12 0 12 0 0.0% +0.000
|
|
2018 10 0 10 0 0.0% +0.000
|
|
2019 149 13 112 16 19.5% +0.022
|
|
2020 157 10 117 19 18.5% +0.042
|
|
2021 201 17 154 23 19.9% +0.020
|
|
2022 188 17 129 29 24.5% +0.034
|
|
2023 194 13 150 19 16.5% +0.018
|
|
2024 226 11 164 40 22.6% +0.066
|
|
2025 191 17 117 38 28.8% +0.062
|
|
2026 138 22 71 28 36.2% +0.004
|
|
(relevant = |xau| >= 0.25, irrel = |xau| <= 0.10, the rest is the grey zone)
|