Evolution of Alpha Signals in Asset Pricing: A Literature Review

Introduction

In asset pricing, “alpha” traditionally refers to abnormal returns beyond what is explained by risk factors. In quantitative trading, however, an alpha signal is any predictive pattern used to forecast excess returns. Academics call them “anomalies” or factors, while practitioners simply call them alphas. These signals range from simple price trends to complex combinations of market data and code. Asset pricing anomalies are trading signals that predict future abnormal returns, violating the Efficient Market Hypothesis. Over the decades, researchers have documented hundreds of such predictors, and quantitative firms have built strategies to exploit them. This review traces the evolution of alpha signals—emphasizing price- and volume-derived signals (e.g., WorldQuant’s “101 Formulaic Alphas”)—and also covers fundamental, macroeconomic, and machine-learning-based alphas. We include both academic findings and practitioner insights up to 2024, outlining how these alphas are constructed and how they have performed, highlighting key developments in each era.

Early Foundations: CAPM and the First Anomalies (1960s–1980s)

The 1960s saw the introduction of the Capital Asset Pricing Model (CAPM), which defined alpha as a risk-adjusted excess return (Jensen’s alpha). Under CAPM, market beta was supposed to explain returns, leaving no persistent alpha. However, by the 1970s and 1980s, cracks appeared in this theory. One of the earliest anomalies was the size effect: small-cap stocks tended to outperform large-cap stocks on a risk-adjusted basis (Banz 1981). Fama and French (1992) confirmed that market beta does not predict stock returns, whereas size and value metrics do. The value effect—stocks with high book-to-market (“value” stocks) outperforming low book-to-market (“growth”) stocks—was demonstrated in the 1980s (Rosenberg et al. 1985) and cemented by Fama-French. Lakonishok, Shleifer, and Vishny (1994) showed value stocks yielded about 10% higher annual returns than glamour stocks, attributing this to investor behavioral biases rather than risk. These findings challenged the notion that markets price securities solely on risk, suggesting the existence of exploitable alpha signals based on firm characteristics.

At the same time, technical analysis—using patterns in price and volume—was long dismissed by academics, but some early studies gave it credence. For example, Brock, Lakonishok, and LeBaron (1992) tested simple moving-average and trading-range break rules on over 90 years of Dow Jones data. They found strong support for these technical strategies: buy signals yielded significantly higher returns (and lower volatility) than sell signals, inconsistent with a random walk or standard models. Returns after technical “sell” signals were negative on average, a result hard to explain by classical theory. This hinted that price trends and patterns could predict future returns—a violation of market efficiency. By the late 1980s, both fundamental-driven anomalies (like size and value) and price-driven signals (like technical trends) were recognized as sources of alpha.

The 1990s: The Rise of Momentum and Price-Based Alphas

The 1990s saw an explosion of research into cross-sectional return predictors, especially those based on past price performance. The most prominent discovery was momentum. Jegadeesh and Titman (1993) showed that stocks which had outperformed over the past 6–12 months continued to outperform recent losers over the next 6–12 months by a wide margin—roughly 1% per month in abnormal return. This “momentum effect” proved robust and pervasive across markets and asset classes. Fama and French later dubbed momentum the “premier anomaly” given its strong and persistent profits. In practice, a standard momentum strategy buys the top decile of stocks by prior 12-month return and shorts the bottom decile, often skipping the most recent month to avoid short-term reversal. Such momentum portfolios earned significant positive alphas in equities and even in bonds, commodities, and currencies (Asness et al. 2013 showed “momentum everywhere”). By the end of the ’90s, momentum had joined size and value as a canonical factor in asset pricing (e.g., Carhart’s 1997 four-factor model added momentum to the Fama-French factors).

Contrarian effects were also noted. Short-term reversal is essentially the opposite of momentum at a one-month horizon: stocks that drop sharply in one month tend to bounce back the next month, and vice versa (Jegadeesh, 1990; Lehmann, 1990). This could reflect liquidity provision or overreaction being corrected. Meanwhile, long-term reversal (De Bondt & Thaler, 1985) found that over 3–5 year periods, past winners underperform past losers—markets overreact in the long run. Momentum appears to flip to reversal over longer horizons: winners keep winning for up to a year, but after 3–5 years they often lag, suggesting initial underreaction and delayed overreaction.

The 1990s also brought renewed interest in volume-related signals. Researchers examined whether trading volume could augment price-based strategies. For instance, Lee and Swaminathan (2000) documented a “momentum life cycle” where incorporating volume trends helped identify when momentum would continue or mean-revert. High trading volume in past winners can indicate sustained momentum, whereas low-volume winners were more prone to reversal. Other studies found abnormal volume itself can be predictive—e.g., spikes in trading volume sometimes signal informed trading or impending news that leads to price drift (Gervais, Kaniel, and Mingelgrin 2001). Though not as extensively published as price anomalies, these insights showed volume is an important dimension of market signals.

Another major anomaly recognized in this era was post-earnings announcement drift (PEAD). Research by Ball and Brown (1968) and Bernard and Thomas (1989) found that after a company announces earnings, the stock’s subsequent returns drift in the direction of the earnings surprise for weeks or even months. If a firm reported earnings much higher than expected, its stock tended to “drift” upward beyond the immediate reaction, yielding a predictable alpha; likewise, negative surprises led to continued underperformance. This was puzzling under efficient markets and indicated a slow incorporation of information. PEAD is a fundamentally-driven alpha signal (based on earnings news), but like momentum it deals with underreaction—just that the information is fundamentals-based rather than purely price.

By the end of the 1990s, dozens of return predictors had been identified. Academic attention began to shift from finding the next anomaly to understanding why these anomalies exist and persist. Two broad schools of thought emerged: risk-based explanations (these “alphas” are compensation for some risk not captured by CAPM) versus behavioral mispricing (investors systematically underreact or overreact, creating profit opportunities). Regardless of cause, the 1990s firmly established that numerous price- and volume-based signals (momentum, reversals, volume/turnover patterns) and event-based or fundamental signals (earnings surprises, etc.) could generate abnormal returns.

The 2000s: Proliferation of Factors and Alpha Signals

The 2000s witnessed a proliferation of documented anomalies—so many that scholars began speaking of a “factor zoo.” Scores of papers identified new predictive stock characteristics: earnings quality indicators, growth and investment metrics, corporate events, and more. An influential example is the accrual anomaly. Sloan (1996) found that companies with high accruals (earnings relying more on accounting accruals than actual cash flow) subsequently earn abnormally low stock returns, compared to firms with low accruals. Firms with aggressively high accruals tend to underperform those with low accruals—consistent with investors initially overvaluing “earnings” not backed by cash. Similarly, researchers discovered a host of other accounting-driven alphas: free cash flow yields, earnings persistence metrics, and analyst forecast errors all showed predictive power for returns in various studies.

Another important development was identifying quality and profitability as drivers of returns. Novy-Marx (2013) showed that gross profitability (gross profits-to-assets) had predictive power comparable to book-to-market in explaining stock returns. Profitable firms earn higher future returns than unprofitable ones, even though one might expect the opposite if prices fully reflected their strength. Adding a profitability factor dramatically improved asset pricing models, explaining many “earnings-related anomalies.” This led Fama and French (2015) to expand their famed model to five factors: they added RMW (robust minus weak profitability) and CMA (conservative minus aggressive investment) to the classic market, size, and value factors. The investment anomaly (CMA) refers to the finding that firms which aggressively expand their assets tend to underperform those that restrain asset growth (Titman, Wei & Xie 2004; Cooper, Gulen & Schill 2008). Thus, low investment (more financially conservative firms) is rewarded by higher returns, which the five-factor model tries to capture.

Meanwhile, practitioners were not idle. Quantitative hedge funds like Renaissance Technologies, D.E. Shaw, Two Sigma, AQR, and others thrived by discovering and trading these alphas. Many signals remained proprietary and secret. It became known anecdotally that top quant funds combined hundreds or thousands of small alphas to create a diversified strategy. For example, Renaissance’s Medallion Fund was reported to use vast numbers of predictive signals, often derived from short-term price patterns, exploiting market microstructure inefficiencies. The success of such funds in the 2000s demonstrated that alpha signals—especially shorter-horizon and higher-turnover ones—could be tremendously profitable if executed well.

On the academic side, by 2010 the count of published return predictors had skyrocketed. Harvey, Liu, and Zhu (2016) documented over 300 factors proposed in the literature. This led to concerns of data mining—perhaps many reported anomalies were false positives. Harvey and colleagues famously suggested raising the statistical hurdle (t-value) required to declare a new factor significant, given the “zoo” of factors. At the same time, efforts were made to consolidate and understand the redundancy among factors. For instance, Feng, Giglio, and Xiu (2020) examined a large set of anomalies and concluded that many of the hundreds of factors are likely redundant, boiling down to a few underlying sources of return. Often, multiple published “anomalies” are just slight variations of one another (e.g., many flavors of value: book-to-market, earnings-to-price, cash-flow-to-price, etc., which are all correlated signals for “cheapness”).

Empirical performance of these factors in the 2000s remained generally strong, though there were notable episodes of crowding. In August 2007, many quant equity strategies simultaneously suffered steep losses in a few days—the “Quant Quake.” It was later hypothesized that overcrowded trades (particularly long/short equity factor trades like momentum and value) unwound rapidly. This highlighted that once an alpha becomes widely known and traded, its returns can diminish or even invert in the short run. McLean and Pontiff (2016) found that after academic publication of an anomaly, its average returns drop by about 30% out-of-sample and 60% post-publication, implying arbitrage capital rushes in to exploit it. Still, many anomalies remain profitable (albeit weaker) even after publication, suggesting limits to arbitrage or continual sources of alpha.

By the late 2000s, new themes emerged. One was low-volatility and low-beta anomalies. Researchers found that the CAPM prediction—higher beta stocks should earn higher returns—had never materialized; in fact, the opposite held: low-beta (or low-volatility) stocks delivered higher risk-adjusted returns than high-beta stocks on average. This was initially observed by Black (1972) and later quantified across markets (Ang et al. 2006). Frazzini and Pedersen (2014) formally constructed a “Betting Against Beta” (BAB) factor that goes long low-beta assets and shorts high-beta assets, yielding significant positive alphas. The existence of a low-volatility premium spurred new fund products (e.g., minimum volatility ETFs) and was interpreted as either due to leverage constraints (many investors can’t leverage low-risk assets, so they over-demand high-beta stocks) or behavioral preferences.

Another theme was “quality” investing. Asness, Frazzini, and Pedersen (2014) defined a composite “Quality Minus Junk” (QMJ) factor, which longs stocks that are high quality (stable earnings, high profitability, low leverage, etc.) and shorts those with “junk” characteristics. They found QMJ earns significant risk-adjusted returns globally. Essentially, the market was not fully pricing in quality differences. This tied together anomalies like profitability, earnings stability, and safety into one broad concept. Quality can be seen as a fundamentally-driven alpha signal complementary to value and momentum.

By the end of the 2000s, investors could access many of these well-known factors through smart beta or hedge fund products. Major asset managers (e.g., AQR, DFA, BlackRock) were packaging factors like value, momentum, quality, size, and low-volatility into investable funds. What had started as academic “alphas” were becoming mainstream—which ironically raised concerns that their returns could be arbitraged away. Nonetheless, niche and undisclosed signals remained the bread and butter of proprietary trading firms.

Early 2010s: WorldQuant’s 101 Alphas and the Big Data Era

Around the mid-2010s, the quant community got an unprecedented peek behind the curtain of a top quantitative hedge fund’s alphas. WorldQuant, an equity quant fund founded by Igor Tulchinsky, published details on some of its proprietary signals. Kakushadze and Tulchinsky (2015) released a study analyzing the statistical properties of 4,000 “real-life” alphas being used in trading. These were largely short-term equity signals with holding periods ranging from intra-day to ~2-3 weeks. They found that the returns of these alphas did not significantly depend on their turnover. In other words, higher-frequency alphas were not inherently less profitable than lower-turnover alphas, after adjusting for costs. They also noted an approximate scaling relationship between an alpha’s volatility and its returns, suggesting diminishing marginal volatility contribution as more alphas are combined. This “4,000 alphas” study was one of the first public confirmations that quant funds were running thousands of signals in parallel and that these signals were only modestly correlated on average (median pairwise correlation was low, ~0.15).

In 2016, “101 Formulaic Alphas”—a white paper by Kakushadze (2016)—listed 101 alpha formulas used at WorldQuant. These formulas, given in pseudo-code, allowed others to recreate them. Most of the 101 alphas are based on price and volume data (daily OHLC prices, returns, volume, VWAP). Many resemble sophisticated technical indicators or cross-sectional signals capturing short-term mean reversion, momentum, or volatility relationships. Some alphas incorporate simple fundamentals or industry classifications to neutralize sector effects. By and large, these are “formulaic” signals built from market data. The 101 alphas had average holding periods of 0.6–6.4 days, annualized Sharpe ratios in a modest range, and low mutual correlations. The real power comes from combining them into a diversified portfolio—a “mega-alpha.” By aggregating hundreds of uncorrelated or lightly correlated signals, a quant fund achieves a high Sharpe ratio portfolio with more stability than any single alpha. Tulchinsky (2015) described this as internal “alpha harvesting” with automatic crossing of trades and continuous rebalancing.

The WorldQuant publications validated that price-volume signals like those in technical analysis (but often more complex and computed across universes of stocks) are heavily used in modern quant trading. They also underscored the arms-race nature of alphas—there are “hundreds of thousands or even millions” of potential alphas being tested with big data and computing power. Each alpha might be very slight, but the combination leads to a formidable edge.

On the academic side, the early 2010s saw broader acceptance that many anomalies are small in isolation and require data-mining and machine learning techniques to identify and aggregate. Researchers started using large-scale datasets of firm characteristics to predict returns. For example, Freyberger, Neuhierl, and Weber (2020) applied LASSO (a machine learning variable selection) to hundreds of stock characteristics to determine which had genuine predictive power. They found that a small subset of characteristics (including momentum, volatility, and liquidity-related measures) explained the majority of cross-sectional return variation, aligning with the notion that many signals are redundant. Most alphas boil down to a few broad types (trend/momentum, mean-reversion, liquidity/volume, value, etc.) plus countless variations in how they’re measured.

Academic-practitioner collaboration increased in this era. Firms like AQR published white papers blending academic rigor with industry data. AQR’s researchers wrote on time-series momentum (Moskowitz, Ooi & Pedersen, 2012), confirming that a trend-following strategy across futures (equities, bonds, commodities, FX) produces high abnormal returns with low correlation to traditional assets. They also explored style premia in alternative contexts (e.g., value and momentum in bonds or across countries). Other industry players like JPMorgan and Goldman Sachs began issuing “quantitative insights” reports, often referencing the factor research in academic literature but applying it to current markets.

By the mid-2010s, alternative data as a source of alpha also emerged. Hedge funds started mining data from news, social media, satellite images, web searches, etc., to find new predictive signals beyond price and fundamental data. For example, news sentiment (e.g., fraction of negative words in news articles) predicts short-term stock returns (Tetlock 2007). Analysis of Twitter or Google Trends data suggested that spikes in search or social media activity could anticipate market moves (Bollen et al. 2011; Da, Engelberg & Gao 2011). These alternative alpha signals gained popularity among quantitative hedge funds, although their success is harder to evaluate publicly. They represent the frontier of alpha sourcing: looking beyond traditional financial metrics to glean any information edge.

Late 2010s to 2020s: Machine Learning and the Maturing Factor Zoo

As we approached the 2020s, two major trends characterized alpha research: machine learning (ML) techniques and a re-examination of the myriad anomalies discovered earlier. A milestone study by Gu, Kelly, and Xiu (2020) systematically applied modern ML algorithms to predict stock returns. They gathered a large set of 94 predictive variables (including technical indicators, fundamental ratios, macro indicators, etc.) for thousands of stocks, and then used methods like trees, random forests, and neural networks to forecast returns out-of-sample. The results were striking: ML forecasts significantly improved performance, in some cases doubling the Sharpe ratio of leading linear strategies. The best-performing techniques (nonlinear models like boosted trees and deep neural nets) capitalized on nonlinear interactions among predictors that classical models miss. For example, an ML model might learn that momentum works best on liquid mid-cap stocks but not on small illiquid ones—a nuanced conditional relationship that a linear factor model would ignore. This led to notably better risk-adjusted returns. Yet, an intriguing finding was that despite scanning through complex interactions, the ML methods largely agreed on the importance of the same major signals: momentum, liquidity, and volatility. Even sophisticated algorithms rediscovered that past return trends, trading volume/liquidity measures, and volatility-related measures are among the most powerful return predictors. This lends credence to the idea that the key sources of equity alpha are relatively few (carry, value, momentum, quality, volatility, etc.), though there are endless ways to combine and implement them.

Machine learning also enabled the use of new data sources and unstructured data. Techniques like natural language processing (NLP) were used to turn textual information (earnings call transcripts, news articles, social media posts) into sentiment scores or thematic signals for trading. Some funds trained ML models on satellite image data (e.g., parking lot traffic at retail stores, crop health from satellite for commodity trading) and claimed to extract alpha from such alternative data. By 2024, “alternative data alpha” had become a buzzword, though rigorous evidence of its broad efficacy is still limited in academic literature. Nonetheless, case studies abound—from using credit card transaction data to predict retailers’ earnings, to scraping job postings data to infer company growth. These novel signals often fall under fundamental or macro categories but are discovered via ML on large datasets, rather than traditional theory.

On the academic side, the late 2010s also saw attempts to consolidate and rationalize the factor zoo. Rather than continue adding new anomalies, researchers asked if we could explain them with a smaller set of “super factors” or risk-based models. Hou, Xue, and Zhang (2015) proposed a “q-factor model” (market, size, investment, and profitability) and argued it could subsume many anomalies related to those themes. Others (Stambaugh & Yuan, 2017) categorized anomalies into those driven by mispricing (e.g., a “misvaluation factor”). Still, anomalies that defy easy categorization—like momentum—continue to challenge any unified risk model.

Another interesting line of research has been anomaly timing and factor momentum. Researchers found that factors themselves exhibit momentum—e.g., a strategy of rotating into factors that have recently performed well and out of those that performed poorly has some success (factor momentum). However, a 2020s re-examination (Gupta and Kelly, 2019; Fan et al., 2022) indicated factor momentum may be weaker than initially thought, and often subsumed by industry momentum effects. Nonetheless, acknowledging that factor returns are time-varying led to the idea of timing signals: e.g., valuation spreads between high and low momentum stocks can signal when momentum is likely to crash (as in the 2009 rebound). AQR researchers and others wrote about “style timing” as an alpha strategy on top of static factor exposures.

The performance of traditional factors in the 2010s also taught humility. Notably, the value factor (HML) suffered an extended drawdown in the late 2010s, underperforming for nearly a decade (2007–2019). This led to vigorous debate in the community. Some argued value was “dead” or had structural headwinds; others (e.g., Asness 2019) argued that value would recover as the spread between cheap and expensive stocks was at extreme levels. Indeed, starting 2020–2022, value made a comeback. The lesson was that even well-established alpha signals go through prolonged cycles of underperformance, requiring conviction (and sometimes capital) to stick with them.

By 2024, the concept of alpha signals had broadened considerably. It includes: classic cross-sectional factors (value, momentum, quality, etc.), style premia across asset classes (e.g., commodity momentum, bond carry), high-frequency signals (order book imbalances, intraday patterns exploited by HFT firms), alternative data signals (sentiment, ESG metrics, web data), and purely statistical patterns discovered via machine learning. Academic publications alongside industry white papers (from firms like AQR, BlackRock, JP Morgan Quant) provide an increasingly unified view: Many purported “new” alphas reduce to known patterns, and successful strategies often combine many signals to diversify away idiosyncratic risk of any one anomaly.

It is also acknowledged that alpha is harder to come by as more players learn about signals. The marketplace for alpha is highly competitive. As one practitioner analogy goes, alphas are like “arbitrage opportunities”—once discovered and traded, they shrink. Yet, human creativity and changing markets continue to yield new signals. For example, the rise of cryptocurrency markets in the 2020s opened a new playing field for alpha search, with researchers examining whether traditional factors like momentum or value apply to crypto assets (preliminary findings show momentum works in crypto too, though with unique twists).

Summary Table of Selected Alpha Signals

The table below summarizes individual alpha signals and factors discussed, along with their source, type, and key notes on construction/performance:

Alpha Signal	Source (Publication/Firm)	Type	Year	Key Construction / Performance Notes
Market (CAPM) Alpha	Jensen (1968)	Theoretical Benchmark	1960s	Alpha = excess return above CAPM. Early studies set the stage for measuring abnormal returns. No persistent alpha if markets are efficient.
Size Effect (SMB)	Banz (1981); Fama & French (1992)	Fundamental	1981/1992	Small-cap stocks outperform large-caps on average (SMB factor). Early anomaly showing market beta wasn’t sole driver. Incorporated in F-F 3-factor model.
Value Effect (HML)	Rosenberg (1985); Fama & French (1992)	Fundamental	1985/1992	High book-to-market (value) stocks earn higher returns than low B/M (growth) stocks, ~+10% annually on avg. Contrarian strategy exploiting investor pessimism about cheap stocks. Core factor in F-F models.
Momentum (Cross-Sectional)	Jegadeesh & Titman (1993); Carhart (1997)	Price-based	1993	Stocks that rose in past 6–12 months continue to outperform losers by ~1%/month on average. “Winners” minus “losers” (UMD factor) is a premier anomaly, pervasive across markets. Requires periodic rebalancing (monthly/quarterly).
Short-Term Reversal	Jegadeesh (1990); Lehmann (1990)	Price-based	1990	1-month contrarian effect: last month’s worst performers tend to outperform the best next month (and vice versa). Likely driven by liquidity provision and short-term overreactions. Often implemented as short-term mean-reversion trades.
Long-Term Reversal	De Bondt & Thaler (1985)	Price-based	1985	3–5 year overreaction: past multi-year winners underperform losers in subsequent years. Implies initial overshooting of prices is corrected over long horizons (behavioral overreaction). Contrarian value-like strategy.
Post-Earnings Announcement Drift (PEAD)	Ball & Brown (1968); Bernard & Thomas (1989)	Event/Fundamental	1968/1989	Stocks continue to drift upward (downward) for weeks/months after very good (bad) earnings news. Traders can long positive surprise firms and short negative surprises to capture abnormal drift. Reflects slow information diffusion.
Accruals Anomaly	Sloan (1996)	Fundamental	1996	Firms with high accruals (aggressive accounting, low earnings quality) earn significantly lower future returns than firms with low accruals. Strategy: long low-accrual (high quality earnings) firms, short high-accrual firms. Exploits earnings quality mispricing.
Profitability (Quality) Factor	Novy-Marx (2013); Fama & French (2015)	Fundamental	2013/2015	More profitable firms earn higher returns, controlling for other factors. Gross profitability/earnings-to-assets has power comparable to book-to-market. Incorporated as RMW (robust minus weak profitability) in F-F 5-factor model. Tied to “quality” – profitable, stable firms have alpha because price doesn’t fully reflect their strength.
Investment Factor (CMA)	Cooper et al. (2008); Fama & French (2015)	Fundamental	2008/2015	Firms that invest conservatively (low asset growth) outperform those with high growth capex or expansions. “CMA” (conservative minus aggressive investment) became a factor in F-F 5-factor model. Captures a broad class of anomalies related to corporate investment and over-extrapolation (e.g. asset growth, financing, IPO/SEO underperformance).
Low-Volatility / Low-Beta	Black (1972); Ang et al. (2006); Frazzini & Pedersen (2014)	Price-based (Risk Anomaly)	1972/2006/2014	Empirically, low-beta (low-vol) stocks have matched or exceeded returns of high-beta stocks, despite lower risk. The BAB (Betting Against Beta) factor (long low-beta, short high-beta, leverage the long side) produces positive risk-adjusted returns. Challenges CAPM; attributed to leverage constraints and investor preferences.
Quality Minus Junk (QMJ)	Asness, Frazzini, Pedersen (2014)	Fundamental (Quality)	2014	A composite “quality” factor: long high-quality stocks (safe, profitable, growing, well-managed) and short low-quality “junk” stocks. Earned significant abnormal returns across global markets. Quality measured via metrics like profitability, stability, low leverage, etc. Markets underprice quality at times, leading to this alpha.
Liquidity Factor	Amihud (2002); Pastor & Stambaugh (2003)	Market Microstructure	2003	Illiquid stocks (high bid-ask spreads or price impact) tend to have higher average returns as compensation for liquidity risk. Pastor-Stambaugh’s liquidity factor and Amihud’s illiquidity measure show a liquidity premium. Also, changes in market-wide liquidity are priced. Strategies exploiting liquidity often require long-horizon (as illiquid stocks can be costly to trade).
Technical Momentum (Moving Avg rules)	Brock, Lakonishok, LeBaron (1992)	Price/Volume (Technical)	1992	Simple technical trading rules (e.g. moving-average crossovers, trading-range breakouts) were shown to have predictive power on long-run Dow Jones data. “Buy” signals (e.g. price breaking above moving avg) yielded higher subsequent returns than “sell” signals. Indicates price trends contain information not captured by equilibrium models. Forms basis of many trend-following strategies.
Momentum (Time-Series)	Moskowitz, Ooi & Pedersen (2012)	Price-based (Macro Trend)	2012	Trend-following within each asset: an asset’s own past 12-month excess return predicts its next 1–12 month return (positive continuation). Implemented by going long assets with positive recent trend and short those with negative trend. A diversified time-series momentum portfolio across asset classes (futures on equities, bonds, FX, commodities) produced high Sharpe and low correlation to stocks/bonds. This is essentially momentum applied in absolute terms (also called “trend” or “CTA” strategy).
Carry Trade (FX Carry)	e.g. Lustig & Verdelhan (2007); Burnside et al. (2011)	Macro (FX Interest Rate)	2007/2011	Carry = “borrow low, lend high”. In currencies, going long high-interest-rate currencies and shorting low-rate currencies has yielded positive returns (the carry trade). The FX carry premium is well-documented, with a high Sharpe ratio historically, though punctuated by occasional crash events (when high-yielding currencies crash during crises). Not explained by CAPM; some explain via exposure to global liquidity risk. Carry is also observed in other asset classes (e.g. bond yield curve roll-down, commodity forward curves).
4,000 Quant Alphas	Kakushadze & Tulchinsky (2015)	Multi-category (Quant Equity)	2015	Study of 4,000 actual alphas (short-term quant trading signals for U.S. stocks, holding 0.7–19 days) from WorldQuant. Found a scaling relation between alpha return and volatility, and importantly no significant performance decay with higher turnover – i.e. even very high-frequency alphas can be as effective as lower-frequency ones. Demonstrated that modern quant strategies employ thousands of alphas concurrently.
WorldQuant 101 Formulaic Alphas	Kakushadze (2016, WorldQuant)	Price/Volume (Quant Equity)	2015/2016	A set of 101 proprietary trading signals (formulas) revealed by WorldQuant. Mostly price-volume based (OHLC prices, returns, volume, volatility, etc.) with a few incorporating fundamentals or industry factors. Examples include cross-sectional rank correlations of returns and volume, mean reversion factors, and short-term momentum signals. These alphas have holding periods ~days and low pairwise correlations. Used in combination to form a larger strategy. Gave insight into practical quant signal construction.
Machine Learning Alphas	Gu, Kelly, Xiu (2020); Various industry implementations	ML / Mixed Data	2020	Application of machine learning to a large “feature library” of signals. Gu et al. showed that tree- and neural network-based models using ~94 characteristics roughly doubled the Sharpe of a traditional linear factor strategy. ML methods captured nonlinear combos (interactions) and dynamics in data. Interestingly, they found the most important features in ML models were familiar ones – variations of momentum, volatility, liquidity, etc. – but ML could time and weight them better. In practice, hedge funds started using ML to find new patterns (e.g. anomaly in combinations of fundamental metrics, or latent factors). ML alphas often involve complex black-box models and require robust backtesting to avoid overfitting.
Alternative Data Signals (Sentiment, ESG, etc.)	e.g. Tetlock (2007); Bollen (2011); Many industry sources	News/Text, Satellite, Other	2007–2020s	New information sources leveraged for alpha: News sentiment (negative/positive tone in news or social media) predicts short-term returns – high pessimism in news can signal undervalued stocks that subsequently rise. Web search trends (Google Trends data) have been used to anticipate attention-driven price moves. Social media sentiment (Twitter feeds) sometimes predicts market volatility or turning points (with mixed success). Satellite imagery & geolocation data used to estimate economic activity (e.g. car counts at retailers, night-time lights for GDP growth) offer macro/stock signals. These alternative data signals are often pursued by quantitative firms to gain an edge, though they require sophisticated processing (ML/NLP) and may be less persistent.
Factor Combination & Timing (“Mega-Alphas”)	WorldQuant (2010s); Feng et al. (2020); Others	Multi-factor	2010s–2020s	The practice of combining many signals into one portfolio – effectively creating a “mega-alpha”. WorldQuant internal approach was to trade a diverse portfolio of hundreds of alphas rather than any single alpha. Benefits include internal crossing of trades (reducing costs) and diversification against signal failure. Academically, this relates to the understanding that many factors are redundant and a few principal components can explain the rest. Also includes factor timing (dynamically weighting factors based on conditions) as an overlay to enhance alpha. The evolution here underscores that the combination of alphas (and avoiding over-crowded ones) is as important as individual signals.

Sources: The table above draws from academic journals (e.g., Journal of Finance, Journal of Financial Economics), working papers, and industry white papers. Key references include Fama & French (1992), Jegadeesh & Titman (1993), Sloan (1996), Novy-Marx (2013), Frazzini & Pedersen (2014), Asness et al. (2014), Moskowitz et al. (2012), Kakushadze & Tulchinsky (2015), Kakushadze (2016), and Gu et al. (2020), among others, as detailed in the text. Each source provided evidence on the existence, methodology, or performance of the respective alpha signals.

Conclusion

Over the past several decades, the concept of alpha in asset pricing has transformed from a simple CAPM intercept into a rich tapestry of alpha signals extracted from every conceivable aspect of market data. What started as a handful of cross-sectional anomalies (size, value, momentum) grew into a “zoo” of hundreds of factors, though many overlap in nature. Price and volume-based alphas—from momentum trends to intricate short-term trading signals—have proven especially central in quantitative strategies, as exemplified by WorldQuant’s formulaic alphas and the enduring success of momentum/trend-following strategies. At the same time, fundamental signals (value, quality, growth, etc.) remain vital, often providing a longer-horizon complement to technical alphas. Macro and cross-asset signals like carry and time-series momentum highlight that alpha is not confined to stock-picking, but is present across markets due to broad risk premia and behavioral trends.

In recent years, the incorporation of machine learning and alternative data has both expanded the frontier of alpha discovery and reinforced the wisdom that many new signals ultimately rediscover the old patterns (momentum, value, carry, etc.) in new guises. The challenge for investors and researchers is twofold: (1) Distill which signals are truly distinct and persistent (separating genuine alphas from data-mined noise), and (2) Combine and adapt these signals in a robust portfolio that can survive changing market regimes and competition. The literature shows progress on both fronts—from higher statistical standards for new factors to sophisticated multi-alpha portfolio construction methods.

In conclusion, alpha signals in asset pricing have evolved through continuous feedback between academia and industry. Academic research identifies anomalies and tests explanations, while quant practitioners implement signals at scale and find practical limitations (like trading costs, crowding effects, turnover constraints). The state-of-the-art by 2024 is a synthesis: multi-factor, multi-asset, machine-assisted strategies that draw on decades of research. Yet, markets are not static—as strategies become crowded, their alpha can decay, prompting the search for new signals. Thus, the evolution of alpha is an ongoing story. History suggests that while the fundamental drivers (value vs. growth, human behavior, risk appetites) remain, the expressions of alpha will continue to adapt with technology and market innovation. By understanding the literature—from the first anomalies to the latest machine learning models—investors and scholars can better appreciate how far we’ve come in decoding alpha, and how far we have yet to go in the ever-competitive quest for outperformance.