File size: 13,512 Bytes
59859a9 | 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 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 | ---
library_name: transformers
tags:
- finance
- sentiment-analysis
- text-classification
- market-impact
- gated-fusion
- multitask-learning
- event-study
datasets:
- NickyNicky/finance-financialmodelingprep-stock-news-sentiments-rss-feed
- siddharthmb/stocks-ohlcv
license: mit
---
# FinImpact Direction1D V4
This repository contains a custom Hugging Face/PyTorch model for financial text impact modeling.
It combines a financial language encoder with market context features through a gated fusion head.
This v4 checkpoint is trained primarily as a 1-day market-impact signal model and adds return-aware auxiliary losses for directional-score alignment and confidence calibration.
## Task
Input:
- Financial news title/body
- Symbol-level market context available before the event date
Outputs:
- `direction_1d`: bearish / neutral / bullish future 1-day abnormal market reaction
- auxiliary diagnostics: `sentiment`, `direction_5d`, and `volatility` heads are still exported, but their default loss weights are zero in this v4 run.
## Architecture
- Encoder: `microsoft/deberta-v3-large`
- Pooling: attention-mask-aware mean pooling over `last_hidden_state`
- Text branch: LayerNorm -> Linear -> GELU -> Dropout
- Numeric branch: LayerNorm -> Linear -> GELU -> Dropout
- Fusion: learned sigmoid gate between text and numeric branches
- Heads: 1D direction classification head plus auxiliary diagnostic heads
- Return-aware training: `P(bullish) - P(bearish)` is softly aligned to realized 1-day abnormal return sign/magnitude
This is intentionally not an embedding-only classifier. The model fine-tunes a contextual encoder and learns a supervised 1-day market-impact boundary.
## Training Data
News source:
- `NickyNicky/finance-financialmodelingprep-stock-news-sentiments-rss-feed`
- Fields used: `symbol`, `publishedDate`, `title`, `text`, `sentiment`, `sentimentScore`
Price source:
- `siddharthmb/stocks-ohlcv`
- Daily close data used to build lagged market features and future-return labels.
Training rows after join: `13930`
Time split:
- train: `10865`
- validation: `1532`
- test: `1533`
Symbols:
`AAPL, ABNB, ADBE, ADI, ADP, AMAT, AMD, AMZN, ARKK, ASML, AVGO, BA, BABA, BAC, BIDU, BMY, CAT, CMCSA, COIN, COST, CRM, CSCO, CVX, DHR, DIS, ELF, GE, GM, GOOGL, HD, IBM, INTC, IOVA, JNJ, JPM, KO, LULU, MA, MCD, MDB, META, MRNA, MSFT, MU, NFLX, NKE, NVDA, ORCL, PEP, PFE, PLTR, PYPL, QCOM, RCL, SHOP, SNOW, SPOT, TGT, TSLA, TSM, TXN, UAL, UBER, UNH, UPS, WMT, XOM, ZM`
## Numeric Features
- `source_sentiment_score`
- `sentiment_abs_score`
- `site_benzinga`
- `site_globenewswire`
- `site_yahoo_finance`
- `site_marketwatch`
- `text_char_len`
- `title_char_len`
- `is_after_close`
- `is_premarket`
- `is_market_hours`
- `published_hour_sin`
- `published_hour_cos`
- `ret_1d_lag`
- `ret_5d_lag`
- `ret_20d_lag`
- `vol_5d`
- `vol_20d`
- `vol_60d`
- `vol_ratio_5_20`
- `drawdown_20d`
- `volume_z_20d`
- `sector_code`
- `market_ret_1d_lag`
- `market_ret_5d_lag`
- `sector_ret_1d_lag`
- `sector_ret_5d_lag`
- `beta_market_60d`
- `beta_sector_60d`
All numeric features are fit/scaled on the training split only.
## Label Construction
- `sentiment` comes from the source dataset label.
- `direction_1d` is generated from future 1-day beta-adjusted abnormal log return after the event anchor date.
- `direction_5d` is generated from future 5-day beta-adjusted abnormal log return.
- A return is neutral when it falls inside a volatility-adjusted threshold.
- `volatility` is based on train-split tertiles of future absolute 5-day return.
Event anchoring shifts after-close UTC news to the next calendar day and uses the next available trading date.
Duplicate same-symbol same-day event titles are removed before training.
## Test Metrics
```json
{
"sentiment": {
"accuracy": 0.009784735812133072,
"macro_precision": 0.0032615786040443573,
"macro_recall": 0.3333333333333333,
"macro_f1": 0.006459948320413436,
"confusion_matrix": [
[
0,
110,
0
],
[
0,
15,
0
],
[
0,
1408,
0
]
],
"predicted_class_counts": {
"bearish": 0,
"neutral": 1533,
"bullish": 0
},
"true_class_counts": {
"bearish": 110,
"neutral": 15,
"bullish": 1408
},
"expected_calibration_error": 0.8247767172447622,
"brier_score": 1.5716914553389048
},
"direction_1d": {
"accuracy": 0.3333333333333333,
"macro_precision": 0.2797514006239241,
"macro_recall": 0.30693224852151285,
"macro_f1": 0.28541099148120913,
"confusion_matrix": [
[
31,
217,
212
],
[
54,
138,
244
],
[
124,
171,
342
]
],
"predicted_class_counts": {
"bearish": 209,
"neutral": 526,
"bullish": 798
},
"true_class_counts": {
"bearish": 460,
"neutral": 436,
"bullish": 637
},
"expected_calibration_error": 0.03496117605077468,
"brier_score": 0.6723385830557309
},
"direction_5d": {
"accuracy": 0.45857795172863663,
"macro_precision": 0.28120269843193446,
"macro_recall": 0.33741824434028733,
"macro_f1": 0.30491929241175314,
"confusion_matrix": [
[
0,
123,
143
],
[
0,
553,
311
],
[
0,
253,
150
]
],
"predicted_class_counts": {
"bearish": 0,
"neutral": 929,
"bullish": 604
},
"true_class_counts": {
"bearish": 266,
"neutral": 864,
"bullish": 403
},
"expected_calibration_error": 0.024249568105288305,
"brier_score": 0.633983250156182
},
"volatility": {
"accuracy": 0.39921722113502933,
"macro_precision": 0.13307240704500978,
"macro_recall": 0.3333333333333333,
"macro_f1": 0.1902097902097902,
"confusion_matrix": [
[
0,
534,
0
],
[
0,
612,
0
],
[
0,
387,
0
]
],
"predicted_class_counts": {
"lower_vol": 0,
"normal_vol": 1533,
"higher_vol": 0
},
"true_class_counts": {
"lower_vol": 534,
"normal_vol": 612,
"higher_vol": 387
},
"expected_calibration_error": 0.1760233100613036,
"brier_score": 0.712886643106662
},
"event_backtest": {
"split": "test",
"transaction_cost_bps": 10.0,
"active_1d_fraction": 0.6568819308545336,
"hit_ratio_1d_active": 0.5789473684210527,
"avg_signal_return_1d_all": 0.0023507131510749455,
"avg_signal_return_1d_all_net": 0.0016938312202204116,
"avg_signal_return_1d_active": 0.0035785931088360393,
"avg_signal_return_1d_active_net": 0.002578593108836038,
"avg_abnormal_signal_return_1d_all": 0.0007162017627715418,
"avg_abnormal_signal_return_1d_all_net": 5.93198319170076e-05,
"avg_abnormal_signal_return_1d_active": 0.0010903051661656142,
"avg_abnormal_signal_return_1d_active_net": 9.030516616561321e-05,
"soft_signal_backtest_1d": {
"mean_abs_signal": 0.049029190093278885,
"active_fraction_abs_score_ge_0.10": 0.11545988258317025,
"hit_ratio_all_gross": 0.532941943900848,
"hit_ratio_active_gross": 0.4463276836158192,
"avg_signal_return_all": -4.1903094825102016e-05,
"avg_signal_return_all_net": -9.093229164136574e-05,
"avg_abnormal_signal_return_all": 6.6492548285168596e-06,
"avg_abnormal_signal_return_all_net": -4.237994289724156e-05
},
"active_5d_fraction": 0.39399869536855836,
"hit_ratio_5d_active": 0.3708609271523179,
"avg_signal_return_5d_all": -0.004762698527096634,
"avg_signal_return_5d_active": -0.012088107354369436,
"confidence_backtests_1d": {
"top_10pct_directional_confidence": {
"num_events": 153,
"mean_confidence": 0.40472039580345154,
"hit_ratio": 0.48366013071895425,
"avg_signal_return": -0.0017291086948829782,
"avg_signal_return_net": -0.002729108694882979,
"median_signal_return": -0.0011537353275343776,
"avg_abnormal_signal_return": 0.0023775517351216227,
"avg_abnormal_signal_return_net": 0.001377551735121622
},
"top_20pct_directional_confidence": {
"num_events": 307,
"mean_confidence": 0.39327916502952576,
"hit_ratio": 0.46579804560260585,
"avg_signal_return": -0.0018346647964980064,
"avg_signal_return_net": -0.0028346647964980075,
"median_signal_return": -0.001494303229264915,
"avg_abnormal_signal_return": 0.00013626769358741198,
"avg_abnormal_signal_return_net": -0.0008637323064125889
},
"top_30pct_directional_confidence": {
"num_events": 460,
"mean_confidence": 0.3858344554901123,
"hit_ratio": 0.5021739130434782,
"avg_signal_return": 0.000168503345380684,
"avg_signal_return_net": -0.0008314966546193169,
"median_signal_return": 0.00047751690726727247,
"avg_abnormal_signal_return": -0.0006967808645482443,
"avg_abnormal_signal_return_net": -0.0016967808645482452
},
"threshold_0.34": {
"num_events": 1108,
"coverage": 0.7227658186562296,
"hit_ratio": 0.5523465703971119,
"avg_signal_return": 0.002400788538323842,
"avg_signal_return_net": 0.001400788538323841,
"median_signal_return": 0.0029742957558482885,
"avg_abnormal_signal_return": 0.00043566977963306825,
"avg_abnormal_signal_return_net": -0.0005643302203669326
},
"threshold_0.36": {
"num_events": 636,
"coverage": 0.41487279843444225,
"hit_ratio": 0.5471698113207547,
"avg_signal_return": 0.001725182487273244,
"avg_signal_return_net": 0.000725182487273243,
"median_signal_return": 0.0027859483379870653,
"avg_abnormal_signal_return": -0.0001395135721791607,
"avg_abnormal_signal_return_net": -0.0011395135721791617
},
"threshold_0.38": {
"num_events": 247,
"coverage": 0.16112198303979125,
"hit_ratio": 0.44129554655870445,
"avg_signal_return": -0.0028902063559107334,
"avg_signal_return_net": -0.0038902063559107343,
"median_signal_return": -0.0022296553943306208,
"avg_abnormal_signal_return": 0.0008437850360328761,
"avg_abnormal_signal_return_net": -0.00015621496396712474
},
"threshold_0.40": {
"num_events": 85,
"coverage": 0.055446836268754074,
"hit_ratio": 0.43529411764705883,
"avg_signal_return": -0.0034284473146887168,
"avg_signal_return_net": -0.004428447314688717,
"median_signal_return": -0.0031941309571266174,
"avg_abnormal_signal_return": 0.0008126156158087884,
"avg_abnormal_signal_return_net": -0.00018738438419121242
},
"threshold_0.42": {
"num_events": 20,
"coverage": 0.01304631441617743,
"hit_ratio": 0.25,
"avg_signal_return": -0.008969856356270612,
"avg_signal_return_net": -0.009969856356270613,
"median_signal_return": -0.011959618888795376,
"avg_abnormal_signal_return": -0.0006033775032847188,
"avg_abnormal_signal_return_net": -0.0016033775032847197
},
"threshold_0.45": {
"num_events": 2,
"coverage": 0.001304631441617743,
"hit_ratio": 0.5,
"avg_signal_return": 0.0010717622935771942,
"avg_signal_return_net": 7.176229357719333e-05,
"median_signal_return": 0.0010717622935771942,
"avg_abnormal_signal_return": 0.01302456425037235,
"avg_abnormal_signal_return_net": 0.01202456425037235
},
"threshold_0.50": {
"num_events": 0,
"coverage": 0.0,
"hit_ratio": null,
"avg_signal_return": null,
"avg_signal_return_net": null,
"median_signal_return": null,
"avg_abnormal_signal_return": null,
"avg_abnormal_signal_return_net": null
}
},
"mean_text_gate_weight": 0.506599485874176,
"std_text_gate_weight": 0.0016385371563956141
}
}
```
The `confidence_backtests_1d` section ranks events by directional confidence, where directional confidence is `max(P(bearish), P(bullish))`.
This is the preferred way to inspect whether the model is useful as a signal filter.
V4 also reports cost-adjusted backtest fields using `10.0` bps per active directional event.
## Important Limitations
This is a research model, not trading advice.
Known limitations:
- Public news sentiment labels can be noisy.
- Daily OHLCV alignment is an approximation; intraday timestamp alignment would be better.
- Market-direction labels are derived from future returns and are sensitive to threshold choice.
- Results should be evaluated out-of-sample by date and by ticker before any practical use.
## Files
- `pytorch_model.bin`: custom gated-fusion model weights.
- `training_config.json`: training and dataset configuration.
- `feature_schema.json`: numeric feature scaler and schema.
- `label_mapping.json`: task label names.
- `metrics.json`: train/validation/test metrics.
- `test_predictions.csv`: event-level test predictions and returns.
- `top_10pct_directional_confidence_1d` in `test_predictions.csv`: marks the highest-confidence directional event subset.
- `confusion_matrix_*.csv`: confusion matrices per split/task.
|