REVIEW 4 major objections 4 minor 1 references
Interdependency between the Stock Market and Financial News
T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read This paper claims that, for the North Korea-related and Tesla stocks studied, stock prices have a larger impact on financial-news sentiment than news sentiment has on stock prices.
desk verdict A sensible research question undermined by look-ahead bias and post-selection; the directional conclusion doesn't follow from the data. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central machinery is a mirrored pair of prediction tasks on the same daily time series. In article-to-stock, the independent variables are the change in sentiment and the number of articles, and the target is the stock price change; in stock-to-article, the independent variables are the price change and trading volume, and the target is the sentiment change. Both tasks are scored by the correlation between predicted and actual outcomes across five models. The sentiment score itself is produced by counting positive words minus negative words in adjective-noun candidates extracted from the top 25 keyword-filtered articles per day, after TF-IDF term weighting. The asymmetry between the two directions is what carries the argument.
What would settle it
Run the same five-model comparison on a rebuilt dataset with all available articles for the same dates rather than the top 25, or apply a Granger-causality test to the daily series: the paper's claim predicts that stock prices at date t help predict sentiment at date t+1 while sentiment at t does not help predict prices at t+1. If the reverse pattern appears, the directional conclusion fails.
Extended reading notes
Core claim
On its own terms, the discovery is a directional asymmetry in time-series prediction. The authors built a daily sentiment index from news articles about North Korea and Tesla, used five learning algorithms (Gaussian processes, linear regression, multilayer perceptron, support-vector regression, and random forest) to test both article-to-stock and stock-to-article prediction, and found that the stock-to-article direction consistently gave higher correlations, especially in classification. In the Tesla classification results, predicting sentiment direction from prices and volume reached an average correlation of 0.826, while predicting price direction from sentiment had a negative average. For North Korea the corresponding numbers were 0.455 and -0.077. The paper therefore concludes that stock prices have a greater impact on news sentiment than news sentiment has on stock prices, and that stock prices respond to social issues before the articles do.
Load-bearing premise
The analysis stands or falls on the assumption that counting positive and negative words in the top 25 keyword-filtered articles per day yields a sentiment score that faithfully represents the news content relevant to these stocks.
Editorial extensions
If this is right
- News-sentiment-based stock prediction is not supported for volatile, event-driven stocks; the paper finds sentiment does not reliably predict the direction of next-day price moves.
- Price movements and trading volume are useful for forecasting near-future news tone, so market data can be treated as an early signal of what will be reported.
- Including past and future article data in the feature set can make stock prediction worse, so simply adding more news history is not a safe improvement.
- Article count and trading volume help regression models but matter little for classifying direction, so their role depends on the prediction task.
Reading between the lines
- An untested implication of the asymmetry is that prices aggregate dispersed private information faster than reporting can; if so, the gap should widen for unexpected events and shrink for scheduled announcements, which could be checked with an event-study design.
- A natural extension is to rerun the same comparison on a broad market index over a longer period with Granger-causality or lead-lag tests; if the asymmetry persists, news-sentiment indices would serve as forecasts of reporting rather than of markets.
- Because the sentiment measure was chosen after comparing its correlation with prices (0.129 vs 0.018), a replication that selects sentiment methods on a training period only would test how much of the asymmetry is an artifact of that selection.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper analyzes the relationship between stock price changes and financial news sentiment for North Korea-related stocks and Tesla, Inc. over roughly three months in 2019. Using daily stock prices and sentiment scores computed from keyword-filtered news articles, the authors apply five machine learning models (GPs, linear regression, MLP, SMOreg, random forest) in both regression and classification settings, and report correlations between predicted and actual change rates. From these correlations, shown in Table 1, the paper concludes that stock prices have a bigger impact on news contents than news has on stock prices, and that stock prices respond to social issues before news articles do.
Significance. If the central claim were supported, the finding would be noteworthy because it contradicts the common assumption that financial news drives stock prices, and it would suggest that prices incorporate information before published news. The paper is also commendable for using a transparent and reproducible data-collection pipeline, multiple classifiers, and public APIs. However, the reported correlations do not constitute a valid test of directional influence, the preprocessing choices are selected on the same data used for evaluation, and no uncertainty quantification is provided. As presented, the evidence is too weak to support the abstract's conclusion.
major comments (4)
- [Dataset Generation and Machine Learning; Table 1] The X ranges include both past and future values: [base date - 5 days, base date + 1 day]. For stockToArticle, X includes stock prices one day after the sentiment date; for articleToStock, X includes articles one day after the stock date. These experiments therefore contain look-ahead information and cannot support a directional or lead-lag conclusion. The only clean experiments are the "base date" rows, which are contemporaneous associations, not evidence that one series moves before the other. The claim that stock prices 'responded before' news requires a proper lagged cross-correlation or Granger-style analysis, which is absent.
- [Candidate Extraction and Synonym Detection; Important Term Identification and Sentiment Analysis; Dataset Generation…] Several key methodological choices are selected based on in-sample correlation with the same stock-price data: the cosine similarity measure is chosen because it yielded 0.129 versus -0.068 for WordNet; the UIC sentiment list is chosen because it yielded 0.129 versus 0.018 for UIC+Stanford; and the X-window [base-5, base+1] is chosen because it yielded the highest average correlation (0.129, versus -0.016 and -0.093). No separate validation set or nested cross-validation is used. Consequently, the correlations reported in Table 1 are post-selection, in-sample statistics, and the apparent advantage of stockToArticle over articleToStock may be an artifact of tuning to the same data.
- [Results and Analysis] The test set consists of the final approximately 34% of a roughly three-month dataset, i.e., around 22 trading days per asset. No confidence intervals, significance tests, or effect-size measures are reported for the correlations in Table 1. Many reported averages (e.g., 0.129, -0.016, 0.014) are close to zero and would not be statistically distinguishable from noise at this sample size. Without uncertainty quantification, the claim that stock prices have a 'bigger impact' than news is not supported.
- [News and Stock Data Collection] The one-day shift added to North Korea article dates is an arbitrary alignment choice that directly affects the computed correlations and any implied causal ordering. The manuscript does not test whether the results are robust to this shift or to alternative alignments. Since the directional claim depends on timing, this unvalidated alignment is a load-bearing assumption.
minor comments (4)
- [Preprocessing] The sentence 'we filtered out an article if the keyword was mentioned less than once' is imprecise; presumably the intended meaning is that articles with zero keyword mentions were filtered out.
- [Preprocessing] The paper states that articles with Levenshtein similarity above 80% were removed, but Levenshtein distance is not a normalized similarity measure; a more precise description of the similarity definition would improve reproducibility.
- [Concluding Remarks] The statement 'stock prices responded to social issues before the articles do' is a stronger causal-temporal claim than the experiments can support; it would be better phrased as a hypothesis for future work.
- [Figure 2] Figure 2 is described as comparing stock price and sentiment index variation, but the caption and text do not indicate the time axis or units clearly; adding axis labels and a legend would improve readability.
Circularity Check
Preprocessing choices are selected on the same stock-price/sentiment correlation that Table 1 then reports, making the reported correlations partly in-sample; the directional claim still retains some independent content.
-
fitted input called prediction
[Section 3.3 (Candidate Extraction and Synonym Detection)]
"The average correlations of WordNet similarity and cosine similarity were –0.068 and 0.129, respectively. Because it had a higher correlation, the second method was used to process synonyms."
The similarity measure used to build the candidate-term representation is chosen because it maximizes correlation with the same stock-price series that later serves as the comparison target in Table 1. The stockToArticle correlations are then computed with that chosen representation on the same data, so they are not an independent measurement of price-to-news influence. The selection criterion and the reported outcome are the same kind of quantity, and the same value 0.129 reappears as the combined regression average.
-
fitted input called prediction
[Section 3.4 (Important Term Identification and Sentiment Analysis)]
"The average correlation of the first method was calculated and found to be 0.129, whereas that of the second method was 0.018. Hence, we used the first method to calculate the sentiment of the candidates."
The sentiment-scoring method is selected because it has the highest correlation with stock-price changes. That same sentiment series is the dependent variable in the stockToArticle experiments and one of the independent variables in the articleToStock experiments. Choosing the sentiment definition by its correlation with the stock series means the subsequent correlations in Table 1 partly measure the success of that selection rather than an independent relationship between news and stocks.
1 more flagged steps
-
fitted input called prediction
[Section 3.5 (Dataset Generation and Machine Learning)]
"The average correlations for the three ranges on combined North Korea and Tesla data for the regression problem were 0.129, –0.016, and –0.093, respectively. Because the first range showed the highest correlation, the rest of the experiments, including the classification, were conducted with the same range of data."
The X-window length is chosen because it produces the highest average regression correlation, and the selected value 0.129 is exactly the combined regression average reported in Table 1 for the 'Past 5 days + base date + future 1 day' range. The window is thus a hyperparameter fitted to the full data, and the reported average is the in-sample optimum of that fitting procedure rather than an out-of-sample estimate of predictive performance.
full rationale
The paper does not rely on self-citation or an imported uniqueness theorem, and no equation in the paper is definitionally identical to its conclusion. The central asymmetry (stockToArticle correlations exceeding articleToStock correlations) is not itself the quantity that was optimized, so the paper is not circular in the strongest sense. However, the pipeline is selected on the same data that later supplies the evidence: the cosine-similarity candidate measure, the UIC sentiment lexicon, and the X-window length are all chosen because they yield the highest correlation with the stock-price/sentiment series, and the chosen X-window's selection value (0.129) is exactly the combined regression average shown in Table 1. Reported correlations are therefore post-selection, in-sample quantities; the stockToArticle results for that window are not an independent test of price-to-news influence. The classification results share the same preprocessing and window, so they do not fully escape the selection. Additionally, the X-range includes a future day, which undermines any lead-lag reading; that is a validity concern rather than a circularity. Overall, the directional conclusion inherits a fitted-input component, but it is not forced by definition, so a moderate score of 4 is appropriate.
Assumptions & free parameters
free parameters (6)
- Input time window length =
base date -5 days to base date +1 day
- Synonym similarity measure =
cosine similarity using word2vec
- Sentiment scoring method =
UIC sentiment word list counting #positive - #negative
- Keyword frequency threshold and article selection =
at least 1 keyword mention for North Korea, at least 3 for Tesla; top 25 articles by keyword frequency
- Levenshtein similarity cutoff =
remove titles more than 80% similar
- North Korea news date shift =
one day added to article dates
assumptions (6)
- domain assumption The daily sentiment score, computed as #positive - #negative on selected candidates, captures the tone of news articles relevant to stock prices.
- domain assumption Pearson correlation between model predictions and actual change rates measures the strength of influence between stock prices and news sentiment.
- domain assumption The five selected North Korea-related stocks represent the North Korean stock market.
- domain assumption The 66%/34% train/test split over roughly three months yields a reliable evaluation.
- domain assumption The observed relationship for North Korea-related stocks and Tesla during April-July 2019 supports a general statement about stock prices and news.
- domain assumption The one-day shift for North Korea aligns US news publication with Korean market dates.
Cite this review
Pith. "Pith review of Interdependency between the Stock Market and Financial News." pith.science (2026). https://pith.science/paper/Q6XAREBJ
@misc{pith2026190900344,
author = {Pith},
title = {Pith review of: Interdependency between the Stock Market and Financial News},
year = {2026},
howpublished = {\url{https://pith.science/paper/Q6XAREBJ}},
note = {Machine review of arXiv:1909.00344}
}
read the original abstract
Stock prices are driven by various factors. In particular, many individual investors who have relatively little financial knowledge rely heavily on the information from news stories when making investment decisions in the stock market. However, these stories may not reflect future stock prices because of the subjectivity in the news; stock prices may instead affect the news contents. This study aims to discover whether it is news or stock prices that have a greater impact on the other. To achieve this, we analyze the relationship between news sentiment and stock prices based on time series analysis using five different classification models. Our experimental results show that stock prices have a bigger impact on the news contents than news does on stock prices.
Figures
Reference graph
Works this paper leans on
-
[1]
Minqing Hu and Bing Liu. 2004. Mining and summarizing cu s- tomer reviews, In Proceedings of the 10th ACM SIGKDD Inte r- national Conference on Knowledge Discovery and Data Mining. Pages 168-177. Joshi Kalyani, H. N. Bharathi, Rao Jyothi. 2016. Stock trend pr e- diction using news sentiment analysis. arXiv:1607.01958 [cs.CL]. K. J. Somaiya College of Engi...
arXiv 2004
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.