Trade the Event: Corporate Events Detection for News-Based Event-Driven Trading

1 Introduction

• Text-feature and sentiment models face poor explainability and low signal-to-noise, while direct rise-or-fall prediction can omit reasons behind price changes.Sentiment can also depend on the author’s standpoint and writing style.Source: §1 · p. 1
• Corporate events are presented as objective facts affecting investor assessments, and token-level detection can identify the article subsequences supporting each event.The paper contrasts this with sentence-level event detection that loses article-wide context and does not indicate event evidence.Source: §1 · pp. 1–2
• The bi-level detector combines token-level event spans with global article context to address the macro-understanding gap of token-only detection.The article-level detector uses low-level predictions together with the article representation.Source: §1 · p. 2
• The strategy makes predictions and trades when an article is published because next-day stock movement may not accurately reflect the article’s influence.The paper motivates this timing by noting that prices may respond immediately to noteworthy news.Source: §1 · p. 2
• The event-driven strategy outperforms baselines in excess returns over the market and average return per transaction, while experiments also indicate news timeliness and corporate events’ usefulness for stock movement.The strategy detects events, identifies supporting subsequences, and trades related stocks.Source: §1 · p. 2
• EDT is introduced as an annotated dataset with more than 300,000 news articles for corporate-event detection and a news-based trading benchmark.The paper also lists the bi-level detector as a contribution integrating macro and fine-grained understanding.Source: §1 · p. 2

2 Problem Definition

• The low-level detector labels each token with one of the predefined event classes or Noevent, yielding a best label sequence for the article.An event is detected at this level when its label appears in the sequence.Source: §2 · p. 2
• A high-level detector uses low-level results to calculate each event’s article-level existence probability, and an event is detected when that probability exceeds a threshold.The final prediction combines both detector levels.Source: §2 · p. 2
• After events are detected, the trading policy determines when to buy or sell the related securities.Source: §2 · p. 2

3 Strategy

• The strategy comprises a bi-level event detector and a trading policy for detecting events in news and trading accordingly.Source: §3 · p. 2

3.1.2 Bi-Level Event Detection

• The bi-level detector classifies tokens for event spans, then combines those predictions with the article representation to estimate event existence at the article level.The low- and high-level losses are summed to optimize both detectors and the text encoder.Source: Figure 2 · p. 3, §3.1.2 · p. 3

3.1.1 Domain Adaptation

• Domain adaptation trains the model with masked-language-model loss on a financial encyclopedia and financial news to address variation in event terminology and meanings.During training, 15% of input tokens are masked and predicted.Source: §3.1.1 · p. 3

3.1.3 Ticker Recognizer

• Ticker recognition matches article text against company–ticker pairs and selects the pair occurring most often when multiple pairs match.The strategy recognizes only one security per article to simplify the setting.Source: §3.1.3 · pp. 3–4

3.2 Trading Policy

• For trading, detected events are mapped by event type to long signals for possible price rises and short signals for possible falls.The strategy trades stocks rather than derivatives to minimize other factors’ influences.Source: §3.2 · p. 4
• Trade-At-End holds a transaction for k trading days, whereas Trade-At-Best closes at the best price within k trading days.Trade-At-Best estimates the maximum profit available within that period for a detected event.Source: §3.2 · p. 4

4 Data

• EDT combines data for corporate-event detection, news-based trading benchmarks, and financial-domain adaptation.The three uses are described in the section overview.Source: §4 · p. 4

4.1.1 Event Type

• The event set covers 11 corporate-event types chosen for relatively predictable and straightforward stock-price impacts, and excludes periodic events such as earnings calls.The authors leave trading on periodic events for future work because investors can access information from multiple sources in advance.Source: §4.1 · p. 4, §4.1.1 · p. 4

4.1.2 Detailed information

• 9721 English articles support event detection: 2266 contain at least one target event and 7455 are non-event articles.The non-event set deliberately includes articles highly similar to event-related ones, such as completion of an announced buyback program.Source: §4.1.2 · pp. 4–5
• Two annotators independently mark token subsequences describing predefined events, resolving disagreements through discussion; the articles come from three newswire sources.The sources are PRNewswire, Businesswire, and GlobeNewswire.Source: §4.1.2 · p. 5

4.2 Data for Strategy Evaluation

• 303893 news articles form the strategy-evaluation corpus, with minute-level timestamps and automatically recognized tickers plus detailed stock-price labels when available.The labels support evaluation by counting price changes for articles identified as trading signals; 106619 evaluation articles have non-empty price labels.Source: §4.2 · p. 5

4.3 Data for Domain Adaptation

• Financial-domain adaptation uses the event-detection training news and 6260 Investopedia documents explaining financial terms and their market roles.Source: §4.3 · p. 5

5.2 Main Results

• Our model outperforms all baselines on average return and excess return under all settings; in 1-day trading, it achieves 1.74% average return and $84443 (844%) excess return.Excess return is total excess over the market, calculated from a $10000 starting amount with $2000 invested per detected signal.Source: §5.2 · p. 7, Table 2 · p. 7
• 98.15% ticker-recognition accuracy was achieved on 1,674 manually labeled articles, though incorrect tickers may impair strategy evaluation.Sentiment-based models indicate stock movements but tend toward small movements and poor average returns; event-based models achieve higher average returns.Source: §5.2 · p. 7
• 9.11% average return under Trade-At-Best marks the trading-signal ceiling and exceeds sentiment-based models.The paper describes this policy as measuring the ceiling of trading signals for each model.Source: §5.2 · p. 7

5.3 Event Analysis

• Reverse Stock Split and Special Dividend are most profitable but rare, while Acquisition, Clinical Trial, and New Contract are common and also move stocks significantly.Trade-At-End yields much lower profits than Trade-At-Best for Acquisition, Clinical Trial, and New Contract, suggesting substantial price oscillations after these events.Source: §5.3 · p. 8
• Similar Trade-At-Best average returns over one and three days for Acquisition, Clinical Trial, and Stock Repurchase indicate prices do not keep changing sharply after day one.The paper suggests strategies could weight events by profitability and choose policies based on their price-change patterns.Source: §5.3 · p. 8

5.4 Profitability in Real-world

• Profitability depends on near-instant execution; trading tens of seconds after publication greatly underperforms the market and yields negative average transaction returns.The analysis assumes negligible time to acquire news and make trading decisions; it also ignores liquidity, which may constrain profitability at larger investment scales.Source: §5.4 · pp. 8–9

• Text-feature methods use news terms or neural representations to predict stock movements, while sentiment and historical stock data are also used (Liu, 2018; Xie and Jiang, 2019).Source: §6 · p. 9
• Event-based stock prediction has used extracted events and event embeddings, while another approach compared event-type and sentiment-type trading signals (Ding et al., 2015; Ben Ami and Feldman, 2017).Source: §6 · p. 9
• Economic-domain event detection has used hand-designed rules and ontologies, followed by data-driven text-classification approaches (Ein-Dor et al., 2019).Source: §6 · p. 9

• Experiments on EDT show the proposed strategy outperforming all baselines, and the results indicate corporate events' timeliness and effectiveness in indicating stock movement.Source: §7 · p. 9
• Future event-detection work will explore external knowledge and few-shot learning for data-imbalance and data-scarce scenarios.Source: §7 · p. 9
• Future trading-policy work will explore more event types and customize policies for each based on potential price-change patterns.Source: §7 · p. 9

• The 3-day trading results are consistent with the results for 1-day and 2-day trading.Source: A Results for 3 days trading · p. 11, Table 6 · p. 11
• Each 3-day transaction starts at the article's publication time and ends after three trading days.Source: Table 6 · p. 11
• The table reports win rate, big-win rate, average return, excess return over the market, and transaction count.Source: Table 6 · p. 11