Event Detection in Financial Markets
Turning noisy news flow into market-moving events - and using those events to model next-day stock direction.
What this project does
Most finance headlines are noise. This system finds the days that look like real market events, extracts the dominant event from the day's headlines, and predicts whether similar events historically pushed the stock up or down.
Identify days with abnormal return behavior (volatility-aware) to focus learning on impactful windows.
Group same-day headlines into clusters and keep the dominant cluster as the event candidate for that day.
Match the new event to similar historical events (semantic similarity) and infer the expected direction using sentiment + neighbors.
Methodology (high-level)
A simple pipeline: find event-like days from price behavior, extract the dominant event from headlines, then predict direction by matching to similar historical events.
Use price time-series to detect days that behave unusually (relative to recent and longer-term volatility). These are likely event days.
On each event day, group similar headlines and keep the dominant cluster as the day's event representation (summarized in the thesis version).
Match the new event to similar historical events and infer the expected return direction using sentiment and historical outcomes.
Technical notes (optional)Toggle
- Event-date detection (EDD): a volatility-aware abnormality score combining short/long horizons and signed/absolute returns.
- Event mining: clustering headlines per day to isolate the dominant event while filtering outliers.
- Thesis-style representation: semantic similarity + automatic cluster summarization for robust matching and cleaner event text.
- Evaluation: standard accuracy on predicted days + a coverage-aware view using RIS.
Tesla walkthrough (visual story)
A single-company view makes the pipeline intuitive: most days have headlines, few days behave like events, and even fewer produce confident predictions.
Compared to the original paper implementation, the thesis version uses semantic similarity and automatic cluster summaries to make the event representation cleaner and matching more robust - while keeping the same core "event mining -> prediction" storyline.
Results snapshot
A quick, portfolio-friendly summary of what improved when we focused training on event dates and event-level representations.
Evaluation (Accuracy + RIS)
Selective prediction changes how you should evaluate a model. I report standard accuracy on predicted days, and also use a coverage-aware metric.
- Directional accuracy computed on the subset of days where the system issues a prediction.
- # Predictions reported alongside accuracy to make the "selective" behavior explicit.
RIS is used to evaluate models that predict on a subset of samples, balancing correctness with selectivity. This makes it easier to compare systems that intentionally abstain under uncertainty.
Reuters key events vs model-detected event headlines (Tesla)
A fun, human-readable check: when Reuters curated a Tesla event timeline, the detected events align closely in meaning - often with completely different wording.
Tesla announces its plan to buy SolarCity, a solar energy system company in which Musk holds a stake, for $2.9 billion.
breaking: tesla makes offer to acquire solar company solarcity
Musk hands over the first Model 3s to employee buyers, announcing over half a million advance reservations for the new electric sedan starting at $35,000. Musk anticipates "at least six months of manufacturing hell."
tesla drops after musk warns of 'manufacturing hell': @randewich $tsla
Tesla pushes back its target to build 5,000 Model 3s per week to the first quarter of 2018 from an original target of December due to production bottlenecks.
tesla reports biggest-ever quarterly loss, model 3 delays $tsla
Musk says Tesla will not need to raise more capital in 2018. Shares jump as much as 6.9 percent.
tesla says no need for capital raise as model 3 output rises
Tesla shares slump after Musk cuts off analysts on a conference call asking about company finances, criticizing their "boring, bonehead" questions. Tesla loses $2 billion in stock market value.
the price of cutting off analysts? for tesla, it's $2 billion
Tesla reports its biggest-ever loss but shares rise on Musk's claims of positive cash flow and profit in the second half of 2018, and signs of more consistent Model 3 production.
happening now: tesla reports second-quarter revenue of $4 billion vs. $2.79 billion reported last year. tesla shares down 1.9 percent in choppy trading after the bell following results $tsla
Musk surprises investors by using Twitter to announce he is considering taking Tesla private at $420 per share, adding "Funding secured."
buyers stampede to tesla stock after elon musk tweets he is considering taking the company private. @alexandriasage reports:
Demo (event matching + explanation)
A recruiter-friendly interaction: paste a headline and see the closest historical events, the cluster summary, and the expected direction.
Where I'd take it next
A few high-leverage upgrades that make the system more robust and more deployable.
Automatically adapt short/long windows (and thresholds) to market regime so "abnormal" stays meaningful across volatility shifts.
Tighten semantic retrieval, and add calibrated confidence so the system can trade off coverage vs reliability explicitly.
Resources
Quick links for readers and reviewers.