q-fin.STOct 15, 2026

Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns

Authors: Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Charishma Takkallapalli, Swapnil Vatsal

Organizations: University of California, Berkeley, Berkeley, California, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Language models that read financial news are usually trained on text from the same years used to test them, so a backtest can reward hindsight. We build news topics from NoLBERT, an encoder pretrained only on text published between 1976 and 1995, adapted into a sentence encoder by contrastive fine-tuning on passages of the same article, with no labels or returns. We compare it with LDA, MiniLM and FinBERT readers on 1.26 million articles from 2012 to 2024, convert each reader's topics into attention shocks and stock exposures, and forecast the returns of the 1,000 largest US stocks with eight models at three rebalancing frequencies. Models are fitted on 2013-2019, choices are made on 2020-2021, and 2022-2024 is held out. The experiment with the best validation score loses money at every frequency. Measuring attention shocks against the past 63 trading days instead of 5 raises the test rank IC of every reader. NoLBERT gives the most stable topics and forecasts about as well as LDA. An equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero before adjusting for the search but not against a price-only benchmark.

Figures & tables

Figure 1: Articles per year, coloured by the split of the return models, and the share of the stock universe with at least one article filed under its ticker in the previous 252 trading days (right axis).
Figure 2: The pipeline. Each article is read by four families of topic readers. A day’s article weights give each topic’s attention, and its deviation from the mean of the previous ww trading days is the attention shock. Narrative exposures, firm news and price characteristics form the features; eight models and their combinations produce forecasts; nine rules turn them into portfolios.
Figure 3: NoLBERT and our adaptation. (a) The pretrained encoder of Kakhbod and Li (2025) : a DeBERTa-v3 base model whose twelve layers combine content and relative-position attention, trained to predict masked words in English text of 1976–1995. (b) Contrastive adaptation: two passages of the same article pass through the shared encoder, are averaged over tokens and normalized, and the symmetric InfoNCE loss rewards placing each passage closest to its partner among the 32 pairs of a batch.
Figure 4: What the adaptation changes. (a, b) Cosine similarity between the headline and the lead of 32 articles from 2022–2024, computed with the settings of the adaptation run; orange dots mark headlines whose most similar lead belongs to another article. Over 512 such pairs, matching pairs have a mean similarity of 0.05 under the base encoder and 0.56 after adaptation, against 0.04 and 0.14 for non-matching pairs. (c) Training loss. (d) Share of passages whose nearest neighbour among the 32 candidates of a batch is their partner, on the 2,048 held-out pairs of the adaptation run and on the 512 pairs from 2022–2024.
BlockNumberDefinition
Price8Market beta and idiosyncratic volatility (252 days); 21-day volatility and maximum return; momentum ( t−251t-251 to t−21t-21 ); 21-day and 5-day reversal; log market value
Narrative exposuresKKEquation ( 4 ) for each topic, with w=5w=5 or w=63w=63
Firm news1+K1+KThe firm’s own articles over 21 days: log⁡(1+count)\log(1+\text{count}) and their share in each topic
Placebo exposuresKKLDA exposures with the shocks moved back by 126 trading days
FinBERT tone8Headline tone ( Araci, 2019 ) over 1, 5, 21 and 252 days, its surprise, article counts and abnormal news volume
Dictionary tone5Loughran–McDonald tone ( Loughran and McDonald, 2011 ) over 1, 5, 21 and 252 days and its surprise
Table 1: Feature blocks. KK is the number of topics of the reader.
ModelSpecificationSettings tried
OLSLeast squares—
Ridgeℓ2\ell_{2} penalty25 penalties
Lassoℓ1\ell_{1} penalty40 penalties
IPCAyi,t+1=zi,t′Γft+1+εy_{i,t+1}=z_{i,t}^{\prime}\Gamma f_{t+1}+\varepsilon , forecast zi,t′Γfˉz_{i,t}^{\prime}\Gamma\bar{f}1, 2, 3 or 5 factors
Sparse IPCAIPCA with a group penalty on the rows of Γ\Gamma1 or 3 factors, 3 penalties
XGBoostHistogram trees, learning rate 0.05, early stopping on validationDepth 1, 2, 3 or 5
Table 2: Forecasting models and the settings tried on the validation period.
Figure 5: Average IC and its tt statistic of the base models by kind of feature set, (a) in the validation period and (b) in the test period. Teal cells are positive, orange negative.
Figure 6: The EBM on price and LDA text. (a) Four single-feature shape functions of the weekly model, for exposures to four LDA topics labelled by their top words. (b) Share of the mean absolute contribution by block. (c) Test Sharpe ratio of the decile long-short portfolio with and without the pairwise terms.
Figure 7: The shock window. (a) Validation IC of the narrative exposures for windows of 1, 5, 21 and 63 trading days, averaged over OLS, ridge and lasso and the six readers. (b–d) Test IC of each reader’s narrative, text and price-and-text experiments with 5-day and 63-day shocks.
Figure 8: The six readers. Coherence: mean NPMI ( Röder et al., 2015 ) of each topic’s ten top words, on text up to 2019 and after. Stability: share of articles that keep their topic when the reader is refitted with another seed. Reliability: split-half reliability of the daily 5-day attention shocks. Evenness: effective number of topics as a share of KK .
Figure 9: Growth of $1 in the buffered long-short portfolios, net of costs: the combinations across readers with 63-day shocks, the price ensemble and the overall pick, with the excess return of the value-weighted market for reference.
  ForecastNet Sharpe ratioBuffered long-short, net
FamilyFeature set ⋅\cdot modelICttDec.Buf.Dec. cwBuf. cwRet. (%)Vol. (%)Max. DDTurn. (%)α\alpha tt
Experiments selected within each family
Overall ⋅\cdot Price + text ⋅\cdot Every set ⋅\cdot LDAPrice + text, LDA ⋅\cdot EBM++ 1.50.53−- 0.23−- 0.370.060.05−- 7.921.2−- 4080−- 0.26
PricePrice ⋅\cdot OLS−- 0.0−- 0.02−- 0.24−- 0.210.230.42−- 3.617.0−- 28117−- 0.80
Text ⋅\cdot NoLBERTExposures, NoLBERT ⋅\cdot IPCA−- 3.5−- 1.48−- 0.62−- 0.82−- 0.36−- 0.36−- 13.516.4−- 4063−- 1.55
Tone/emb.Price + all news ⋅\cdot IPCA++ 2.51.640.941.040.260.3211.410.9−- 13751.54
Table 3: Monthly rebalancing: test-period forecasts and portfolios (36 months).
   ForecastReturnSR Turnover
ComparisonFrequencyPairsΔ\Delta ICttΔ\Delta Ret.ttBaseNewImpr.BaseNew
63-day against 5-day shocksMonthly46++ 1.281.8++ 5.82.20.240.6291106102
 Biweekly46++ 0.872.0++ 4.72.00.050.34808180
 Weekly48++ 0.491.7++ 3.91.6−- 0.240.00858988
Ranked against raw targetMonthly32++ 0.260.5−- 1.0−- 0.40.250.3759102101
 Biweekly34++ 0.952.5++ 3.41.10.060.30797676
Table 4: Paired comparisons. Each pair of experiments differs only in the stated choice. Δ\Delta IC: mean difference in IC; Δ\Delta Ret.: mean difference in net long-short return (percentage points a year); tt : Newey–West tt statistics; SR: median Sharpe ratio without and with the change; Impr.: share of pairs whose Sharpe ratio improves (%); Turnover: median turnover per period (%).
  Against zerovs. price ensemblevs. 5-day shocksvs. overall pick
 Combination, ruleSRttppDSRΔ\DeltappΔ\DeltappΔ\Deltapp
MonthlyPrice + text, decile1.122.360.0450.04++ 0.460.463++ 0.740.149++ 1.350.196
 Price + text, buffered1.192.420.0400.06++ 0.660.267++ 0.850.119++ 1.560.146
 Text only, decile0.861.600.1570.01++ 0.200.836++ 1.310.028++ 1.090.182
 Text only, buffered0.841.770.1220.01++ 0.310.722++ 1.320.025++ 1.210.129
BiweeklyPrice + text, decile0.691.410.1820.00++ 0.590.266++ 0.460.152++ 0.960.101
Table 5: Significance of the 63-day combinations across readers. SR: net Sharpe ratio; tt , pp : Newey–West test of the mean return with fixed- bb pp -value; DSR: deflated Sharpe ratio. Each comparison gives the Sharpe ratio difference Δ\Delta and the pp -value of the Ledoit–Wolf test.
Figure 10: Robustness to the period and to the seed.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1 : Average IC of each model over the 26 feature sets of the main design, in the test period (bars; teal for models fitted on ranked returns) and the validation period (circles).
Figure A2 : Portfolio rules over the base-model experiments. (a) Median and interquartile range of the Sharpe ratio (information ratio against the universe for the long-only rules and the tilt). (b) Median turnover per period.
Figure A3 : The best of all experiments by rule and frequency, in the main design (375 experiments) and with every experiment (617), against the best Sharpe ratio expected without skill. A line is coloured when the best beats that benchmark.
  MonthlyBiweeklyWeekly
Kind of feature setnnICttPos.SRICttPos.SRICttPos.SR
Price8++ 2.22.4750.41++ 1.21.7100−- 0.07++ 1.61.5880.09
Narrative exposures, 5 d48−- 0.5−- 0.9400.03−- 0.4−- 0.731−- 0.20−- 0.3−- 0.740−- 0.46
Narrative exposures, 63 d48++ 1.82.1900.84++ 1.01.5850.39++ 0.61.3830.03
Firm news48−- 0.2−- 0.431−- 0.82++ 0.10.250−- 1.30−- 0.0−- 0.142−- 1.53
Text, 5 d48−- 0.5−- 0.8350.00−- 0.5−- 0.827−- 0.20−- 0.4−- 1.027−- 0.57
Appendix
Table A1 : Test-period forecasts by kind of feature set. IC: average IC of the base-model experiments; tt : tt statistic of the group’s average IC per period; Pos.: share of experiments with a positive IC (%); SR: median Sharpe ratio of their decile long-short portfolios.
 MonthlyBiweeklyWeekly
 ICSRICSRICSR
Reader5 d63 d5 d63 d5 d63 d5 d63 d5 d63 d5 d63 d
LDA++ 0.70++ 2.850.111.20++ 0.20++ 2.010.040.81−- 0.06++ 0.94−- 0.360.26
LDA (all articles)−- 1.30++ 0.55−- 0.85−- 0.13−- 0.88++ 0.41−- 0.79−- 0.13−- 0.60++ 0.45−- 0.94−- 0.26
MiniLM−- 0.44++ 1.34−- 0.020.62−- 0.34++ 1.08−- 0.450.35−- 0.27++ 0.70−- 0.64−- 0.05
FinBERT++ 0.20++ 1.970.200.80−- 0.14++ 0.70−- 0.210.29++ 0.01++ 0.55−- 0.29−- 0.13
Appendix
Table A2 : Test-period forecasts by reader: average IC of the base-model experiments on the reader’s narrative, text and price-and-text sets with 5-day or 63-day shocks, and the median Sharpe ratio (SR) of their decile long-short portfolios.
 MonthlyBiweeklyWeekly
ModelVal.ICttSRVal.ICttSRVal.ICttSR
OLS++ 0.4−- 0.6−- 1.3−- 0.20++ 1.1−- 0.3−- 0.6−- 0.45++ 0.8++ 0.10.2−- 0.61
Ridge−- 0.1−- 0.1−- 0.20.20++ 1.4++ 0.00.0−- 0.15++ 0.9++ 0.20.7−- 0.34
Lasso++ 0.9−- 0.9−- 1.3−- 0.45++ 1.9−- 0.5−- 0.9−- 0.31++ 0.9++ 0.10.3−- 0.41
IPCA++ 0.9++ 0.61.1−- 0.02++ 0.7−- 0.1−- 0.4−- 0.41++ 0.7++ 0.10.4−- 0.44
Sparse IPCA−- 0.1++ 0.51.1−- 0.17++ 0.2++ 0.30.80.06++ 0.4−- 0.7−- 1.1−- 0.55
Appendix
Table A3 : Forecasts by model, averaged over the 26 feature sets of the main design. Val.: validation IC; IC and tt : test IC and its tt statistic; SR: median test Sharpe ratio of the decile long-short portfolios.
 MonthlyBiweeklyWeekly
RuleSRPos.Turn.SRPos.Turn.SRPos.Turn.
Decile long-short0.1657101−- 0.034879−- 0.343076
Buffered long-short0.2357720.105450−- 0.233845
Decile, cap-weighted0.0350114−- 0.203194−- 0.342292
Buffered, cap-weighted0.035181−- 0.163559−- 0.263054
Rank-weighted0.075168−- 0.104451−- 0.332651
Appendix
Table A4 : Portfolio rules over the base-model experiments. SR: median test Sharpe ratio; Pos.: share of experiments with a positive one (%); Turn.: median turnover per period (%).
 MonthlyBiweeklyWeekly
 Main (375)All (617)Main (375)All (617)Main (375)All (617)
RuleBestNo skillSPABestNo skillSPABestNo skillSPABestNo skillSPABestNo skillSPABestNo skillSPA
Decile long-short1.061.290.592.151.370.000.731.330.881.691.400.030.661.320.930.931.420.81
Buffered long-short1.311.270.161.961.360.010.831.320.721.641.390.060.751.310.961.181.400.56
Decile, cap-weighted1.251.340.161.611.410.210.951.400.691.211.490.700.691.470.940.991.570.91
Buffered, cap-weighted1.511.360.021.881.410.030.851.400.951.291.480.560.991.450.691.191.550.73
Appendix
Table A5 : The best of the search. Best: best test Sharpe ratio (information ratio for † ); No skill: best expected without skill, from the bootstrap with every mean return set to zero; SPA: pp -value of Hansen’s test.
  ForecastNet Sharpe ratioBuffered long-short, net
FamilyFeature set ⋅\cdot modelICttDec.Buf.Dec. cwBuf. cwRet. (%)Vol. (%)Max. DDTurn. (%)α\alpha tt
Experiments selected within each family
Overall ⋅\cdot Text ⋅\cdot FinBERTText, FinBERT ⋅\cdot EBM−- 0.5−- 0.45−- 0.27−- 0.13−- 0.76−- 0.26−- 1.612.2−- 2050−- 0.20
PricePrice ⋅\cdot IPCA++ 0.70.450.04−- 0.01−- 0.24−- 0.27−- 0.317.3−- 1995−- 0.12
Price + text ⋅\cdot LDAPrice + text, LDA ⋅\cdot EBM−- 1.2−- 0.64−- 0.52−- 0.63−- 0.76−- 0.95−- 14.322.9−- 4777−- 1.19
Tone/emb.Price + all news ⋅\cdot Lasso++ 0.80.740.070.21−- 0.41−- 0.392.512.2−- 15450.55
Appendix
Table A6 : Biweekly rebalancing: test-period forecasts and portfolios (78 fortnights).
  ForecastNet Sharpe ratioBuffered long-short, net
FamilyFeature set ⋅\cdot modelICttDec.Buf.Dec. cwBuf. cwRet. (%)Vol. (%)Max. DDTurn. (%)α\alpha tt
Experiments selected within each family
Overall ⋅\cdot Price + text ⋅\cdot Every set ⋅\cdot Ranked targets ⋅\cdot LDAPrice + text, LDA ⋅\cdot EBM−- 1.3−- 0.91−- 1.08−- 0.98−- 1.01−- 0.88−- 21.221.6−- 5188−- 1.87
PricePrice ⋅\cdot IPCA++ 3.02.310.430.52−- 0.30−- 0.199.518.3−- 151060.84
Text ⋅\cdot NoLBERTText, NoLBERT ⋅\cdot EBM++ 0.30.35−- 0.08−- 0.04−- 0.160.02−- 0.614.0−- 17300.28
Tone/emb.Price + tone ⋅\cdot IPCA++ 3.02.300.380.46−- 0.18−- 0.148.418.4−- 171050.75
Appendix
Table A7 : Weekly rebalancing: test-period forecasts and portfolios (156 weeks).
 Against zeroAgainst a benchmark
PortfolioSRttppPSRDSRBenchmarkΔ\Delta SRLW ttLW ppBoot. pp
Monthly
All readers, price + text, 63 d, decile1.122.360.0450.950.04price ensemble++ 0.460.820.4630.437
      same, 5-day shocks++ 0.741.640.1490.155
      text alone++ 0.260.550.6260.632
      overall pick++ 1.351.450.1960.159
Appendix
Table A8 : Significance of single portfolios with fixed forecasts, with all benchmarks. SR: Sharpe ratio; tt , pp : Newey–West test of the mean return; PSR and DSR: probabilistic and deflated Sharpe ratios. Against a benchmark: Sharpe ratio difference, Ledoit–Wolf test ( tt , pp ) and paired stationary bootstrap ( pp ).
More in this topic

Financial Forecasting

Financial forecasting uses market, company, and economic data to predict prices, returns, volatility, and business outcomes for decisions and risk management.

See all 88 papers

Explore similar work

CardsList

  1. From Financial Sentiment Classification to Return Predictability: A QLoRA Benchmark of Large Language Models

    Aug 4, 2026Fusheng LuoLLM EvaluationQuantitative Finance

    Johns Hopkins University Baltimore, Maryland, USA

  2. Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI

    Jun 10, 2026Ali M Karaoglu, Shreyank N GowdaFinancial ForecastingZero-Shot Text Classification

    School of Computer Science, University of Nottingham, Nottingham, United Kingdom

  3. DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining

    Mar 12, 2026Yutong Yan, Raphael Tang, Zhenyu Gao +2Look-Ahead Bias in Language ModelsLanguage Model Pretraining

    Department of Finance, CUHK Business School, The Chinese University of Hong Kong · Centre for Artificial Intelligence, University College London