pith. sign in

arxiv: 2504.02429 · v2 · pith:J3UFKIZ3new · submitted 2025-04-03 · 💻 cs.CE

MulFSA: Multi-level Financial Sentiment Analysis Framework for Bond Market

Yiwei Liu , Junbo Wang , Lei Long , Xin Li , Ruiting Ma , Yuankai Wu , Xuebin Chen This is my paper

Pith reviewed 2026-05-22 21:47 UTC · model grok-4.3

classification 💻 cs.CE
keywords financial sentiment analysisbond marketcredit spread forecastingmulti-level sentimentChinese bond textssentiment indexduration-aware smoothing
0
0 comments X

The pith

MulFSA combines firm-level and industry-level sentiments with duration smoothing to cut credit spread forecast errors by over 10 percent.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a Multi-level Financial Sentiment Analysis framework that processes bond market texts at the firm-specific micro level and the industry-specific meso level, then applies duration-aware smoothing to capture how textual signals persist or fade over time. It builds this on a new 1.35 million text corpus covering the Chinese bond market from 2013 to 2023 and derives a daily composite sentiment index. When this index is added to standard credit spread models, the paper reports a 10.25 percent drop in mean absolute error and an 11.94 percent drop in mean absolute percentage error. The approach is motivated by the claim that single-level sentiment methods miss the layered and time-dependent nature of risk information in bond markets.

Core claim

MulFSA integrates firm-specific micro-level sentiment, industry-specific meso-level sentiment, and duration-aware smoothing inside pre-trained and large language models. Applied to the 1.35 million text Chinese bond corpus, it produces a daily composite sentiment index whose inclusion in forecasting models yields 10.25 percent MAE reduction and 11.94 percent MAPE reduction for credit spreads, with shifts aligning to major risk events and firm crises.

What carries the argument

The MulFSA framework that systematically integrates firm-specific micro-level sentiment, industry-specific meso-level sentiment, and duration-aware smoothing to model the latency and persistence of textual impact on bond risk.

If this is right

  • Credit spread forecasting models achieve lower errors when the daily composite sentiment index is included.
  • Sentiment index movements track documented social risk events and firm-specific crises in the Chinese bond market.
  • The multi-level construction works on a 1.35 million text corpus spanning a full decade of market activity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same layered extraction could be tested on equity or derivative markets to check whether multi-level signals add value outside bonds.
  • If the duration smoothing step is the main driver, simpler time-decay adjustments might achieve similar gains without separate micro and meso layers.
  • Daily index values might support real-time monitoring dashboards that flag emerging credit events earlier than price data alone.
  • The framework leaves open whether the improvements hold when the underlying language models are swapped for lighter or open-source alternatives.
  • keywords

Load-bearing premise

The duration-aware smoothing and the separation into micro- and meso-level sentiments each add independent predictive value beyond single-level sentiment, and the 1.35 million texts accurately represent the full Chinese bond market from 2013 to 2023.

What would settle it

Running the same credit-spread forecasting regressions on a new out-of-sample period or market while adding the MulFSA index and finding no MAE or MAPE reduction, or finding that a single-level sentiment index performs equally well.

Figures

Figures reproduced from arXiv: 2504.02429 by Junbo Wang, Lei Long, Ruiting Ma, Xin Li, Xuebin Chen, Yiwei Liu, Yuankai Wu.

Figure 1
Figure 1. Figure 1: Overview of Multi-Level Sentiment Analysis Framework. The composite sentiment extracted from the firm-specific and industry-specific levels is aggregated and then incorporated as an additional feature to downstream tasks. Abstract—Existing financial sentiment analysis methods often fail to capture the multi-faceted nature of risk in bond markets due to their single-level approach and neglect of temporal dy… view at source ↗
Figure 2
Figure 2. Figure 2: Task Decomposition. Each task corresponds to one of our contributions in Section. I. III. METHODOLOGY A. Terminology And Task Decomposition In this work, we define sentiment s of a single text as a continuous value ranging from −1 to 1: s ∈ [−1, 1]. (1) There are financial implications at three polarities: s = −1 indicates pessimism, suggesting deteriorating expectations or higher risk; s = 0 is neutral, m… view at source ↗
Figure 3
Figure 3. Figure 3: After forwarding the text through BERT, we obtain both [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SLSA Procedure. Algorithm 2: RAG Mapping Input: unlabeled D2,β, GPT agent f2(·), embedding model f3(·) Output: meso-level sentiment sβ 1 foreach text j ∈ D2,β do 2 text sentiment = f2(j) ; 3 text embedding = f3(j) ; 4 topics cosine similarity list l = empty list ; 5 foreach topic n ∈ B′ do 6 l append cos(f3(j), f3(n)) ; 7 end 8 top5 nearest topics = filtered(l) ; 9 sβ = the sum of f2(j) multiplied by G and… view at source ↗
Figure 5
Figure 5. Figure 5: The Transformation Relationships. its similarity cj,n. This equation implies that the meso-level sentiment sβ,i,k of an arbitrary bond on a given day is the average of the industry sentiment sβ,m,k relevant to that day. Each industry sentiment sβ,n,j,k is computed as a weighted average of the topic sentiments sβ,n,j,k, where the similarity score cj,n controls the influence strength of different industries … view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of Rolling Window. when k = K is the current time, it integrates historical data for real-time inference, and when k = K is a past date, it backtests. On days with fewer texts, a bond’s sentiment is driven by its industry’s sentiment, indicating lower market attention and a greater reliance on industry factors. However, not every bond or industry has daily text coverage, resulting in sparse t… view at source ↗
Figure 7
Figure 7. Figure 7: Sentiment Heatmap for 40 Industries (2013-2023). Values exceeding −1 and 1 were truncated for display. The marked period exhibits a significant sentiment shift vis-a-vis its preceding period, aligning with the corresponding social event. ` We visualized the industry-specific sentiment matrix Sβ obtained using Qwen2.5-3B-Instruct and Algorithm. 2, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Industry Sentiment Time Series Decomposition of Automobile (2013-2023). From top to bottom, the four subplots represent the origin, trend, seasonality, and residuals. TABLE III: Comparison Results on Forecasting Target q. Forecasting Target t + q Sentiment MAE (e-5) MAPE (e-3) p ∆MAE (%↓) ∆MAPE (%↓) t + 1 11.2366 9.4581 n/a n/a n/a t + 1 " 11.3680 8.3686 0.0812 -1.1693 11.5194 t + 2 8.9683 8.0033 n/a n/a n… view at source ↗
Figure 9
Figure 9. Figure 9: The Correlation Matrix of Features. TABLE IV: Ablation Results. Micro-Level Sentiment Meso-Level Sentiment Duration Function MAE (e-5) MAPE (e-3) p ∆MAE (%↓) ∆MAPE (%↓) 8.9683 8.0033 n/a n/a n/a " 8.7890 10.0594 0.0573 1.9991 -25.6900 " 8.7621 8.0367 0.0472 2.2989 -0.4170 " " 15.567 43.479 ≈ 0.0 -73.579 -443.258 * * 8.9489 8.0205 0.1013 0.2159 -0.2148 " " " 8.6765 7.1257 0.0373 3.2539 10.9658 [PITH_FULL_I… view at source ↗
Figure 10
Figure 10. Figure 10: Component Visualization of CATL. Since industry￾specific sentiment is transmitted from topic sentiment via the Knowledge Graph G, and topic sentiment is aggregated from daily sentiment, if all industries associated with this company exhibit the same polarities on certain days, this can lead to extreme values appearing on those days in the visualization. ations. However, after aggregation and the applicati… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of Composite Sentiment Dynamics of Defaulted Bonds Preceding defaults. The left side of each subplot displays the bond code. “SH” indicates the Shanghai Stock Exchange, and “IB” indicates the Investment Bank. V. CONCLUSION Benefits. In this work, we independently constructed a multi-source heterogeneous dataset and leveraged PLMs and LLMs to extract firm-specific and industry-specific sentim… view at source ↗
read the original abstract

Existing financial sentiment analysis methods often fail to capture the multi-faceted nature of risk in bond markets due to their single-level approach and neglect of temporal dynamics. We propose Multi-level Financial Sentiment Analysis (MulFSA) based on pre-trained language models (PLMs) and large language models (LLMs), a novel framework that systematically integrates firm-specific micro-level sentiment, industry-specific meso-level sentiment, and duration-aware smoothing to model the latency and persistence of textual impact. Applying MulFSA to the comprehensive Chinese bond market corpus constructed by us (2013-2023, 1.35M texts), we extracted a daily composite sentiment index. Empirical results show statistically measurable improvements in credit spread forecasting when incorporating sentiment (10.25% MAE and 11.94% MAPE reduction), with sentiment shifts closely correlating with major social risk events and firm-specific crises. Project Page: https://mulfsa.github.io/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes MulFSA, a framework that integrates firm-specific micro-level sentiment, industry-specific meso-level sentiment, and duration-aware smoothing using PLMs and LLMs to construct a daily composite sentiment index from a 1.35M-text Chinese bond market corpus (2013-2023). It claims this index yields statistically measurable improvements in credit spread forecasting (10.25% MAE reduction, 11.94% MAPE reduction) over a no-sentiment baseline and correlates with major risk events.

Significance. If the multi-level decomposition and smoothing demonstrably add predictive value beyond simpler sentiment aggregation, the work would advance financial sentiment analysis by addressing multi-faceted risks and temporal dynamics in bond markets. The scale of the constructed corpus (1.35M texts) is a clear strength for domain-specific empirical work.

major comments (2)
  1. [Abstract / Empirical results] Abstract and empirical results section: The headline improvements (10.25% MAE, 11.94% MAPE) are reported only versus a no-sentiment baseline. No ablation results are provided for micro-only, meso-only, or unsmoothed single-level variants, leaving open whether the multi-level architecture and duration-aware smoothing contribute independent value as asserted in the abstract.
  2. [Abstract] Abstract: The reported percentage reductions supply no information on the base forecasting regressor, train/test splits, cross-validation procedure, or statistical significance tests (e.g., Diebold-Mariano), which are required to substantiate that the gains are attributable to the MulFSA index rather than confounding factors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important aspects of clarity and empirical rigor that we will address in the revision. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract / Empirical results] Abstract and empirical results section: The headline improvements (10.25% MAE, 11.94% MAPE) are reported only versus a no-sentiment baseline. No ablation results are provided for micro-only, meso-only, or unsmoothed single-level variants, leaving open whether the multi-level architecture and duration-aware smoothing contribute independent value as asserted in the abstract.

    Authors: We agree that ablation experiments are required to isolate the incremental value of the multi-level decomposition and duration-aware smoothing. The current manuscript reports only the full MulFSA versus the no-sentiment baseline. In the revised version we will add a dedicated ablation subsection in the empirical results that compares the complete framework against micro-only, meso-only, and unsmoothed single-level variants, using the same forecasting setup and significance tests. revision: yes

  2. Referee: [Abstract] Abstract: The reported percentage reductions supply no information on the base forecasting regressor, train/test splits, cross-validation procedure, or statistical significance tests (e.g., Diebold-Mariano), which are required to substantiate that the gains are attributable to the MulFSA index rather than confounding factors.

    Authors: The abstract currently omits these methodological specifics. The full manuscript describes the forecasting setup in Section 4, but to make the claims self-contained we will revise the abstract to include a concise statement of the base regressor, the temporal train/test protocol, the cross-validation approach, and the statistical tests employed. The empirical results section will also be expanded with explicit reporting of these elements. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper constructs the MulFSA sentiment index from an external 1.35M-text corpus via PLMs/LLMs with micro/meso levels and duration smoothing, then reports empirical forecasting gains on credit spreads versus a no-sentiment baseline. No equations or steps reduce the index construction or the reported MAE/MAPE improvements to self-definition, fitted parameters reused as predictions, or load-bearing self-citations. The chain relies on independent data processing and model application rather than tautological re-use of inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the framework description implies possible fitted smoothing weights and corpus-construction choices but none are stated.

pith-pipeline@v0.9.0 · 5704 in / 1159 out tokens · 29870 ms · 2026-05-22T21:47:42.545056+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 75 canonical work pages · 4 internal anchors

  1. [1]

    International recessions,

    F. Perri and V . Quadrini, “International recessions,” American Economic Review, vol. 108, no. 4-5, pp. 935–984, 2018

    work page 2018

  2. [2]

    Investor sentiment and the cross-section of stock returns,

    M. Baker et al. , “Investor sentiment and the cross-section of stock returns,” The journal of Finance , vol. 61, no. 4, pp. 1645–1680, 2006

    work page 2006

  3. [3]

    News and asset pricing: A high-frequency anatomy of the sdf,

    S. Aleti and T. Bollerslev, “News and asset pricing: A high-frequency anatomy of the sdf,” The Review of Financial Studies, p. hhae019, 2024

    work page 2024

  4. [4]

    Emotions in macroeconomic news and their impact on the european bond market,

    S. Consoli, L. T. Pezzoli, and E. Tosetti, “Emotions in macroeconomic news and their impact on the european bond market,” Journal of International Money and Finance , vol. 118, p. 102472, 2021

    work page 2021

  5. [5]

    Financial sentiment analysis: Techniques and applica- tions,

    K. Du et al. , “Financial sentiment analysis: Techniques and applica- tions,” ACM Computing Surveys , vol. 56, no. 9, pp. 1–42, 2024

    work page 2024

  6. [6]

    Identifying corporate credit risk sentiments from financial news,

    N. Ahbali, X. Liu, A. Nanda, J. Stark, A. Talukder, and R. P. Khandpur, “Identifying corporate credit risk sentiments from financial news,” in Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, 2022, pp. 362–370

    work page 2022

  7. [7]

    How to gauge investor behavior? a comparison of online investor sentiment measures,

    D. Ballinari and S. Behrendt, “How to gauge investor behavior? a comparison of online investor sentiment measures,” Digital Finance , vol. 3, no. 2, pp. 169–204, 2021

    work page 2021

  8. [8]

    Year and industry-level accounting narrative analysis: Readability and tone variation,

    E. Efretuei, “Year and industry-level accounting narrative analysis: Readability and tone variation,” Journal of Emerging Technologies in Accounting, vol. 18, no. 2, pp. 53–76, 2021

    work page 2021

  9. [9]

    Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis,

    Z. Liu et al., “Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 5487–5496

    work page 2024

  10. [10]

    Chasing noise,

    B. Mendel and A. Shleifer, “Chasing noise,” Journal of Financial Economics, vol. 104, no. 2, pp. 303–320, 2012

    work page 2012

  11. [11]

    FinBERT: Financial Sentiment Analysis with Pre-trained Language Models

    D. Araci, “Finbert: Financial sentiment analysis with pre-trained lan- guage models,” arXiv preprint arXiv:1908.10063 , 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1908

  12. [12]

    Fingpt: Democratizing internet-scale data for financial large language models,

    X.-Y . Liu et al., “Fingpt: Democratizing internet-scale data for financial large language models,” arXiv preprint arXiv:2307.10485 , 2023

    work page arXiv 2023

  13. [13]

    BloombergGPT: A Large Language Model for Finance

    S. Wu, O. Irsoy, S. Lu, V . Dabravolski, M. Dredze, S. Gehrmann, P. Kambadur, D. Rosenberg, and G. Mann, “Bloomberggpt: A large language model for finance,” arXiv preprint arXiv:2303.17564 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Large language models in finance (finllms),

    J. Lee, N. Stevens, and S. C. Han, “Large language models in finance (finllms),” Neural Computing and Applications , pp. 1–15, 2025

    work page 2025

  15. [15]

    A survey on aspect-based sentiment analysis: Tasks, methods, and challenges,

    W. Zhang, X. Li, Y . Deng, L. Bing, and W. Lam, “A survey on aspect-based sentiment analysis: Tasks, methods, and challenges,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 11, pp. 11 019–11 038, 2022

    work page 2022

  16. [16]

    Big data: Deep learning for financial sentiment analysis,

    S. Sohangir, D. Wang, A. Pomeranets, and T. M. Khoshgoftaar, “Big data: Deep learning for financial sentiment analysis,” Journal of Big Data, vol. 5, no. 1, pp. 1–25, 2018

    work page 2018

  17. [17]

    An investigation of investor sentiment and specula- tive bond yield spreads,

    G. Cerci et al. , “An investigation of investor sentiment and specula- tive bond yield spreads,” in International Interdisciplinary Business- Economics Advancement Conference , 2015, p. 224

    work page 2015

  18. [18]

    Good debt or bad debt: Detecting semantic orientations in economic texts,

    P. Malo et al., “Good debt or bad debt: Detecting semantic orientations in economic texts,” Journal of the Association for Information Science and Technology, vol. 65, no. 4, pp. 782–796, 2014

    work page 2014

  19. [19]

    Www’18 open challenge: financial opinion mining and question answering,

    M. Maia, S. Handschuh, A. Freitas, B. Davis, R. McDermott, M. Zarrouk, and A. Balahur, “Www’18 open challenge: financial opinion mining and question answering,” in Companion proceedings of the the web conference 2018 , 2018, pp. 1941–1942

    work page 2018

  20. [20]

    When flue me ets flang: Benchmarks and large pre-trained language model for financial domain

    R. S. Shah, K. Chawla, D. Eidnani, A. Shah, W. Du, S. Chava, N. Raman, C. Smiley, J. Chen, and D. Yang, “When flue meets flang: Benchmarks and large pre-trained language model for financial domain,” arXiv preprint arXiv:2211.00083, 2022

    work page arXiv 2022

  21. [21]

    Combining enterprise knowledge graph and news sentiment analysis for stock price prediction,

    J. Liu, Z. Lu, and W. Du, “Combining enterprise knowledge graph and news sentiment analysis for stock price prediction,”Hawaii International Conference on System Sciences , 2019

    work page 2019

  22. [22]

    Bias propagation in economically linked firms,

    T. Jochem and F. S. Peters, “Bias propagation in economically linked firms,” Available at SSRN 2698365 , 2019

    work page 2019

  23. [23]

    Too sensitive to fail: The impact of sentiment connect- edness on stock price crash risk,

    J. Cao et al. , “Too sensitive to fail: The impact of sentiment connect- edness on stock price crash risk,” Entropy, vol. 27, no. 4, p. 345, 2025

    work page 2025

  24. [24]

    News sentiment and bank credit risk,

    L. A. Smales, “News sentiment and bank credit risk,” Journal of Empirical Finance, vol. 38, pp. 37–61, 2016

    work page 2016

  25. [25]

    The timeliness of the bond market reaction to bad earnings news,

    M. L. Defond and J. Zhang, “The timeliness of the bond market reaction to bad earnings news,” Contemporary Accounting Research , vol. 31, no. 3, pp. 911–936, 2014

    work page 2014

  26. [26]

    A survey of sentiment analysis in social media,

    L. Yue et al. , “A survey of sentiment analysis in social media,” Knowledge and Information Systems , vol. 60, pp. 617–663, 2019

    work page 2019

  27. [27]

    Learning Universal Sentence Representations with Mean-Max Attention Autoencoder

    M. Zhang, Y . Wu, W. Li, and W. Li, “Learning universal sentence representations with mean-max attention autoencoder,” arXiv preprint arXiv:1809.06590, 2018. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  28. [28]

    Large lan- guage models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large lan- guage models are zero-shot reasoners,” Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

    work page 2022

  29. [29]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    work page 1901

  30. [30]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al. , “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems , vol. 35, pp. 24 824–24 837, 2022

    work page 2022

  31. [31]

    Foundations of portfolio theory,

    H. M. Markowitz, “Foundations of portfolio theory,” The journal of finance, vol. 46, no. 2, pp. 469–477, 1991

    work page 1991

  32. [32]

    Framing theory,

    D. Chong and J. N. Druckman, “Framing theory,” Annu. Rev. Polit. Sci., vol. 10, no. 1, pp. 103–126, 2007

    work page 2007

  33. [33]

    Nerlove, D

    M. Nerlove, D. M. Grether, and J. L. Carvalho, Analysis of economic time series: a synthesis . Academic Press, 2014

    work page 2014

  34. [34]

    Credit spreads as predictors of real-time economic activity: a bayesian model-averaging approach,

    J. Faust et al. , “Credit spreads as predictors of real-time economic activity: a bayesian model-averaging approach,” Review of Economics and Statistics, vol. 95, no. 5, pp. 1501–1519, 2013

    work page 2013

  35. [35]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    P. Wang et al., “Qwen2-vl: Enhancing vision-language model’s percep- tion of the world at any resolution,” arXiv preprint arXiv:2409.12191 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    C-pack: Packaged resources to advance general chinese embedding,

    S. Xiao, Z. Liu, P. Zhang, and N. Muennighoff, “C-pack: Packaged resources to advance general chinese embedding,” 2023

    work page 2023

  37. [37]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    DeepSeek-AI, “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” 2025

    work page 2025

  38. [38]

    No contagion, only interdependence: measuring stock market comovements,

    K. J. Forbes and R. Rigobon, “No contagion, only interdependence: measuring stock market comovements,” The journal of Finance, vol. 57, no. 5, pp. 2223–2261, 2002

    work page 2002

  39. [39]

    Clean evidence on peer effects,

    A. Falk and A. Ichino, “Clean evidence on peer effects,” Journal of labor economics, vol. 24, no. 1, pp. 39–57, 2006

    work page 2006

  40. [40]

    Stl: A seasonal-trend decomposition,

    R. B. Cleveland et al. , “Stl: A seasonal-trend decomposition,” J. off. Stat, vol. 6, no. 1, pp. 3–73, 1990

    work page 1990

  41. [41]

    Permutation tests for studying classifier performance

    M. Ojala et al., “Permutation tests for studying classifier performance.” Journal of machine learning research , vol. 11, no. 6, 2010

    work page 2010

  42. [42]

    All models are wrong, but many are useful: Learning a variable’s importance by studying an entire class of prediction models simultaneously,

    A. Fisher, C. Rudin, and F. Dominici, “All models are wrong, but many are useful: Learning a variable’s importance by studying an entire class of prediction models simultaneously,” Journal of Machine Learning Research, vol. 20, no. 177, pp. 1–81, 2019. Yiwei Liu is an undergraduate student expected to obtain a Bachelor’s degree in Software Engineering, Si...

    work page 2019

  43. [43]

    Shibor (Shanghai Interbank Offered Rate) in March

    work page

  44. [44]

    Manufacturing PMI (Purchasing Managers’ Index)

    work page

  45. [45]

    Macroeconomic Prosperity Index: Leading Index

    work page

  46. [46]

    PPI (Producer Price Index): Year-over-Year for the Current Month

    work page

  47. [47]

    GDP (Gross Domestic Product): Year-over- Year for the Current Quarter

    work page

  48. [48]

    CPI (Consumer Price Index): Year-over-Year for the Current Month

    work page

  49. [49]

    Aggregate Financing to the Real Economy (AFRE): Year-over-Year at Period-End

    work page

  50. [50]

    SWS 3 Primary Industry Index Trading Indicator 11

    Yield on Government Bonds (for the Corre- sponding Period) Industrial Indicator 10. SWS 3 Primary Industry Index Trading Indicator 11. Trading V olume Firm Financial and Operational Indicators

    work page

  51. [51]

    Current Assets 14 Firm Financial and Operational Indicators

    work page

  52. [52]

    Non-Current Liabilities

    work page

  53. [53]

    Total Shareholders’ Equity

    work page

  54. [54]

    Cash Flow from Operations

    work page

  55. [55]

    Cash Flow from Investment

    work page

  56. [56]

    Cash Flow from Finance

    work page

  57. [57]

    Debt-to-Asset Ratio (%)

    work page

  58. [58]

    Tangible Net Worth Debt Ratio (%)

    work page

  59. [59]

    Gross Profit Margin (%)

    work page

  60. [60]

    Net Profit Margin (%)

    work page

  61. [61]

    Return on Assets (%)

    work page

  62. [62]

    Operating Profit Margin (%)

    work page

  63. [63]

    Average Return on Equity (%)

    work page

  64. [64]

    Operating Cycle (Days)

    work page

  65. [65]

    Inventory Turnover Ratio

    work page

  66. [66]

    Accounts Receivable Turnover Ratio

    work page

  67. [67]

    Current Asset Turnover Ratio

    work page

  68. [68]

    Shareholders’ Equity Turnover Ratio

    work page

  69. [69]

    Total Asset Turnover Ratio Firm Comprehensive Credit Indicators

    work page

  70. [70]

    Remaining Credit Utilization Ratio

    work page

  71. [71]

    Month-over-Month Change in Credit

    work page

  72. [72]

    Secured Credit Ratio TABLE VIII: Feature Attribution. Feature Name Feature Importance 1: PPI 0.05599892602199476 2: Macroeconomic Climate Index 0.05008474777898123 3: AFRE 0.048741595825980824 4: Yield on Government Bonds 0.04777380830739202 5: Shibor 0.04202742096446867 6: CPI 0.03426918656953126 7: GDP 0.024902612696883587 8: Cash Flow from Finance 0.02...

    work page

  73. [73]

    Pessimistic (-1): The text describes factors reflecting negative market sentiment in macro-financial contexts, particularly the bond market. These include concerns about economic fundamentals (such as slowdowns or rising unemployment), expectations of tighter monetary policy and higher interest rates, reduced market liquidity, and risk events like default...

    work page

  74. [74]

    The content is mostly unrelated to macro financial markets, especially the bond market

    Neutral (0): The text describes neutral market sentiment, characterized by stable economic fundamentals, unchanged monetary policy with priced- in expectations, balanced market liquidity, and absence of major risk events. The content is mostly unrelated to macro financial markets, especially the bond market

    work page

  75. [75]

    It is related to macro financial markets, particularly the bond market

    Optimistic (1): The text reflects positive market sentiment, with mentions of stronger economic fundamentals, expectations of rate cuts or accommodative policy, ample liquidity, and easing risk events such as debt relief or rating upgrades. It is related to macro financial markets, particularly the bond market. **Output Requirements:** - a) First, identif...

    work page 2023