Six common sources for factor mining: financial reports, technical indicators, analyst expectations, news sentiment, macroeconomic data, and sector rotation.

The essence of quantitative factors is to capture a market pattern using a single number. Below, we present 5–6 concrete factors from each of six data sources, along with explanations of why they work. I. Financial Report Factors Extracted from balance sheets, income statements, and cash flow statements. • Quarter-on-quarter revenue growth rate: The simplest and most stable metric. Companies reporting three consecutive quarters of >15% growth are highly likely in an upturn phase. • Surprises in Earnings (SUE): (Actual EPS − Expected EPS) divided by the standard deviation of expected EPS. An SUE > 2 is typically viewed as "explosive earnings," often triggering positive short-term price feedback. • Quarter-over-quarter change in gross margin: An increase in gross margin often signals enhanced pricing power or lower costs—making the *change* more predictive than the absolute level. • Accruals ratio: (Net income − Operating cash flow) / Total assets. A higher ratio indicates lower "cash quality" of earnings and correlates with poorer future returns—a well-documented phenomenon known as the "accruals anomaly." • Cash conversion cycle: Days inventory outstanding + Days sales outstanding − Days payables outstanding. A shorter cycle reflects higher operational efficiency. • Change in long-term debt ratio: A sudden rise in long-term debt / total assets within one quarter may signal expansion—or risk—and must be interpreted in context of industry norms. II. Technical Indicator Factors Derived from price, volume, time, and speed. • Skewness factor: Skewness of daily returns over the past 20 days. Stocks with negative skew (long left tail) historically exhibit higher probabilities of sharp downside moves; backtests show portfolios with low skewness deliver higher returns. • Turnover rate change: Today’s turnover rate divided by its 20-day moving average. A ratio exceeding 2×, coupled with stagnant price, often signals a reversal. • Intraday volatility factor: (Daily high − Daily low) / Prior day’s closing price. Narrowing volatility often precedes breakouts; sudden expansion frequently accompanies trend acceleration or reversals. • Williams %R: Measures where the closing price lies within the recent high–low range. Readings below −80 indicate oversold conditions; above −20, overbought. Yet in A-share markets, overbought conditions often persist—so momentum context is essential. • Deviation from Volume-Weighted Average Price (VWAP): (Current price − VWAP) / VWAP. A large positive deviation accompanied by high volume suggests concentrated buying pressure, often followed by short-term pullbacks. • Closing position factor: Percentage position of the closing price between the day’s high and low. A top-quartile close followed by a gap-down open signals weakness. III. Analyst Expectation Factors Derived from secondary processing of earnings forecasts and rating data. • Magnitude of EPS forecast upward revision: (This month’s consensus EPS forecast − Last month’s) / Last month’s consensus. Revisions exceeding 5% warrant attention—but beware potential manipulation by institutions. • Unexpected rating factor: Convert analysts’ “buy” recommendation ratios into Z-scores. Stocks with excessively high Z-scores tend to underperform subsequently—because broad optimism often leaves little room for alpha. • Change in forecast dispersion: A sudden narrowing of dispersion signals growing consensus, often preceding clear directional moves in price. • Coverage breadth factor: Number of analysts covering the stock over the past three months. A sudden increase often coincides with material corporate events. • Target-price premium: (Average target price / Current price) − 1. Higher premiums do not guarantee higher future returns—but a sudden, synchronized upward revision across analysts is predictive.

Can earnings surprises be predicted in advance? Two approaches—from dataset construction and methodology—are proposed at ACL 2026.

Predicting whether a company’s next-quarter earnings report will trigger a "surprise" (positive or negative) is among the most valuable—and hardest—tasks in quantitative investing. Yet there’s an awkward reality: the historical data needed for this task is either prohibitively expensive proprietary text, or reduced to bare-bones transcripts stripped of critical contextual cues—such as management’s tone during earnings calls or key points highlighted in presentation slides. At ACL 2026, two papers jointly argue that rather than continuing to hit dead ends—either chasing data monopolies or relying solely on plain text—we should instead pursue two parallel paths: fully open-sourcing and multimodalizing the data, while simultaneously stripping away redundant signals inside models. These works push the field forward significantly—one by building better data, the other by redesigning methods. __________________________________________________ I. Why Predicting Earnings Surprises Matters So Much An earnings surprise is simply the difference between a company’s actual reported earnings per share (EPS) and Wall Street analysts’ consensus expectation. Once interpreted by the market as "beat" or "miss," this gap often triggers substantial price swings within just a few days—making it one of the most actionable signals for event-driven strategies. Being able to anticipate the direction of the next surprise thus yields tangible, real-world returns. But the data is extremely hard to obtain. Audio recordings, verbatim transcripts, and presentation slides—the primary raw materials—have long been monopolized by a handful of commercial data vendors. They’re costly, and most research only accesses cleaned-up text versions; audio and slide content remain inaccessible. This directly constrains state-of-the-art large models from performing at their full potential. Worse still, earnings call transcripts routinely exceed 5,000 words and are saturated with boilerplate language and industry jargon—presenting models with long, noisy, highly redundant texts. Extracting truly discriminative signals in such environments has become the field’s central challenge. __________________________________________________ II. Notable Works Worth Reviewing FinCall-Surprise: Open-sourcing multimodal earnings call data to establish a benchmark for surprise prediction First is FinCall-Surprise—a large-scale, open-source, multimodal benchmark dataset for earnings surprise prediction, described in the paper "FinCall-Surprise: A Large-Scale Multi-modal Benchmark for Earning Surprise Prediction." What problem does it address? Prior research on earnings surprise prediction has largely relied on expensive, proprietary, text-only data—leaving advanced models unable to access audio or slide content. The authors target precisely this gap: creating a publicly available dataset that includes synchronized verbatim transcripts, full-length audio recordings, and corresponding presentation slides—truly incorporating the "multimodal" dimension. Their goal is to provide a fair, standardized evaluation arena for future models.

Element Solutions Announces Quarterly Earnings, Beats Expectations by $0.03 EPS

Apr 28, 2026

Element Solutions reported quarterly earnings of $0.41 EPS and $840 million revenue, exceeding analyst estimates, though stock price dipped slightly following the announcement.

MarketBeat

2026-04-28

Element Solutions Inc Reports Q1 Revenue Exceeding Estimates

2026-04-28

Element Solutions: Q1 Earnings Snapshot

2026-04-28

Element Solutions (ESI) Stock Rises on Q1 2026 Earnings

2026-04-28

Element Solutions Announces Quarterly Earnings, Beats Expectations by $0.03 EPS

2026-04-29

Element Solutions Q1 2026 slides: electronics surge drives guidance raise