Who’s Still Earning Alpha? Patient Investors and Sentiment Traders

The return edge from patient capital

TLDR

• Stocks with more long-horizon owners earn higher future returns

• The effect is strongest when you restrict the universe of stocks to those with ‘high idiosyncratic volatility’ or ‘recent losers’

What the authors did

They built a firm-level “Horizon” metric from 13F filings that reflects the ownership-weighted holding period, sorted the stocks on Horizon, and tested future returns. The result: longer-horizon ownership predicts higher returns.

How they constructed the firm-level “Horizon”

They started with quarterly 13F filings and focused on active institutions (indexers were excluded). For each stock each quarter, they calculated how long these active owners have been continuously holding it (in quarters), then took an ownership-weighted average across those owners.

That weighted average is the firm’s “Horizon” for that quarter. Effectively, “how patient are this company’s active shareholders, on average?” Longer average holding periods = higher Horizon

What they showed

They sorted stocks by their Horizon score and compared the subsequent performance of high-Horizon vs low-Horizon names (controlling for standard risk factors).

The basic question was: do firms whose owners hold longer earn higher future returns than peers? The answer is yes, and the horizon premium is bigger exactly where career risk and limits to arbitrage are worst.

“High idiosyncratic-volatility” (idio-vol) stocks and “recent losers” are the spots most active managers hate to sit through, so there is less patient capital competing there and more mispricing for those who can hold. Patient capital was rewarded with an even higher spread in those situations (so high idio-vol stocks with long-term holders did even better compared to short-term holders, i.e., with low-Horizon).

Why high idio-vol

High idio-vol stocks swing a lot for stock-specific reasons. That raises tracking error and career risk for managers, who are judged on short windows and face redemptions after drawdowns.

When holding a volatile name goes against you, clients leave or you get fired. That makes managers avoid or underweight these stocks even when they are cheap. With fewer patient buyers, prices can drift below fundamentals and offer higher expected returns to investors who can hold through noise.

This is the core idea behind limits to arbitrage: idiosyncratic risk is costly because it cannot be hedged, so it deters arbitrage and lets mispricing persist. The myopia paper says the horizon-return link is strongest in stocks that are hard for short-term managers to hold; classic work explains why idio-vol is one of those frictions. See “The Limits of Arbitrage”, March 1997; and “Costly arbitrage and the myth of idiosyncratic risk”, April 2006.

Why recent losers

Recent losers are uncomfortable to own because underperformance begets outflows and scrutiny. Flow-performance sensitivity and termination risk push managers to cut losers and rotate into safer, consensus names, even if that is suboptimal long-term. The result is a scarcity of patient capital on the loser side, so expected returns rise to compensate whoever is willing and able to wait.

The myopia paper finds the horizon spread is larger in recent losers. The career-concerns literature shows why managers have incentives to avoid these positions when their job depends on near-term marks. See “Managerial Career Concern and Mutual Fund Short-Termism”, March 2008; and “Career Concerns of Mutual Fund Managers”, May 1999.

The bottom line

Think of a long–short portfolio that is long the top-Horizon names and short the bottom-Horizon names. That long–short seems to earn a positive return on average.

When you restrict the universe to high idio-vol stocks or to recent losers, the same long–short earns even more. So the premium to holding firms with patient owners is strongest exactly where short-term pressure scares most managers away.

Timing markets with cross-border news

TLDR

• Daily long-short overlays built from news sentiment beat buy-and-hold in 13 of 14 developed markets, with lower betas and smaller drawdowns

• Adding cross-country news linkages can improve performance further

What the authors did

They generated country-level signals by training tree-based models on news sentiment. The output was a probability that a country’s stock market index would go up or down the next day, and they bought (or sold) index futures according to that binary signal. 

The result: Net of trading frictions in the paper’s setup, Sharpe improved meaningfully and drawdowns shrank.

Specifically, across all 14 markets that the authors examined, local-news strategies had a net Sharpe that was 82.5% higher than a buy & hold strategy, and max drawdowns were down by about one-third.

The paper’s U.S. example shows a local strategy Sharpe of 1.34 (1.11 net of trading costs) with net CAPM alpha of 16% per year. These numbers are consistent with a low-beta strategy earning 16% excess return (alpha) with ~12% annualized volatility.  The high alpha indicates very low market correlation, while >1 Sharpe ratio means that it accomplished those returns relatively efficiently (without too bumpy of a ride).

How they generated “country-level signals”

They started from GDELT’s news stream, which is a set of continuously updated datasets and APIs that ingest, translate, and annotate news from around the world, then publish structured records you can query in near real time.

The news stream tags each article with themes, the country it comes from, and the country it talks about. 

This enabled the authors to build daily sentiment features for each market two ways: “local” news about a country from its own media, and “foreign” news about that country from other countries’ media.

Those features then fed a predictive model that output a next-day return signal per country. In short: you can turn millions of tagged articles into daily sentiment numbers for each market, both local and cross-border.

What are “tree-based models” and why do they fit here

The authors used tree ensembles like Random Forests (See “Random Forests”, January 2001). Trees split the data according to rules such as “if U.S. tone ↑ and energy theme ↑, then…,” and ensembles average many such trees.

Trees handle non-linear interactions, mixed feature types, and lots of correlated inputs without much preprocessing, which is exactly what you get with high-dimensional text features. They are also relatively robust and interpretable compared to deep nets for tabular data.

Tree ensembles are powerful machine learning models that combine the predictions of multiple, weaker decision trees to form a single, more accurate, and robust model. Instead of relying on a single decision tree, which can be prone to overfitting and instability, tree ensembles use various techniques like bagging and boosting to aggregate the results from individual trees. 

For example, in Random Forests, many trees are built on different subsets of the data, and their final predictions are averaged for regression or determined by a majority vote for classification.

What was the trading decision

After the model scores tomorrow’s return sign for a country, the authors took a binary stance in that market’s index future: +1 if the expected return is positive, −1 if negative. No leverage ramp, just long or short 1x notional, which keeps beta low and implementation simple. Costs are set to a small, fixed number typical for liquid index futures in academic tests. This makes the timing benefit come from signal quality, not from hidden leverage.

How did foreign news add incremental alpha to local news

When they added features built from foreign media coverage about a country, the model’s out-of-sample performance improved over using only local news. That says information travels across borders and markets react to how other countries’ media frame a story. The U.S. sits at the center of this network, but cross-border links still add distinct predictive content, which shows up as higher alphas and better Sharpe after costs.

Why it matters & how to use it

If you build a lean version in production, start with a local-news overlay per market, then layer in foreign-news features from key trade or financial partners. Using GDELT or a comparable feed gives you the source/target structure you need; tree ensembles are a pragmatic baseline before you try heavier models. Validate every step out-of-sample and keep the trading rule simple so the lift, if any, comes from the signal, not leverage or parameter fiddling.

Illustrative example

• Tuesday market close (time = T): Model with local+foreign news sentiment features outputs that the probability for the market index to go up the next day is 62% ⇒ the trading signal is +1 ⇒ You buy 1x notional of S&P 500 futures.

• Wednesday (T+1): S&P future ends the day +0.40% ⇒ Gross strategy return is +0.40% ⇒ Round-turn trading cost: ~0.04% (2bps in, 2bps out) ⇒ net ≈ 0.36%

• At Wednesday’s market close: Re-run features. Suppose P(up) = 47% ⇒ signal flips to −1. You short 1x notional for Thursday. And so forth…

Where foreign news helps

Suppose local U.S. news is mildly positive (P(up) = 55%), but overnight European and Asian coverage of the U.S. earnings season and policy risk is strongly positive. The cross-border features push the probability to 62%, firming the +1 decision. That “incremental” lift from foreign-to-U.S. links is exactly what the paper documents: adding global linkages raises out-of-sample performance beyond local-only signals.

Disclaimer

This publication is for informational and educational purposes only. It is not investment, legal, tax, or accounting advice, and it is not an offer to buy or sell any security. Investing involves risk, including loss of principal. Past performance does not guarantee future results. Data and opinions are based on sources believed to be reliable, but accuracy and completeness are not guaranteed. You are responsible for your own investment decisions. If you need advice for your situation, consult a qualified professional.