What Actually Predicts Stock Returns? A Machine Learning Perspective

Decades of academic research and institutional quant work have converged on one answer: no single factor reliably predicts stock returns in all market conditions. But certain categories of features — used together, in the right framework — have demonstrated persistent predictive power across long time horizons. Here's what the data actually shows, and how retail traders can use it.

Quant-Builder.ai — Simplifying Quant Trading: Try a Free Demo at quant-builder.ai/learn

Why This Question Is Hard to Answer

Stock return prediction is noisy. Any single feature — even well-researched ones like value or momentum — goes through multi-year periods where it stops working entirely. This is why simple rules like "buy low P/E stocks" or "buy stocks above the 200-day moving average" work sometimes and fail spectacularly other times.

Machine learning doesn't solve this by finding the one perfect feature. It solves it by finding non-linear combinations of many features that together have more predictive power than any individual signal — and by validating those combinations out-of-sample so you know the pattern is real.

Category 1: Valuation

Value factors have been documented since the 1970s. Stocks trading at low multiples of earnings, sales, or book value have historically outperformed over long periods — not every year, but across full market cycles. The intuition is straightforward: you're paying less per dollar of fundamentals.

Key features: P/E ratio, P/S ratio, EV/EBITDA, price-to-book, P/FCF.

The caveat: value has long stretches where it underperforms (2010–2020 was brutal for pure value strategies). It works better as one input among many than as a standalone signal.

Category 2: Earnings Quality and Revision Trends

Earnings surprises and upward analyst revisions have strong short-to-medium-term predictive power. A stock that beats earnings estimates and receives upward revisions tends to continue outperforming — the market is slow to fully reprice new information. This is sometimes called the Post-Earnings Announcement Drift (PEAD) effect.

Key features: EPS surprise, revenue surprise, analyst revision direction, earnings growth rate, margins.

Category 3: Momentum

Price momentum — the tendency of recent winners to keep winning over 3–12 month horizons — is one of the most replicated findings in finance. It persists across geographies, asset classes, and time periods.

Key features: 12-1 month return (excluding the most recent month), relative strength vs. sector, 52-week high proximity.

The catch: momentum crashes hard in sharp reversals. It's powerful but requires careful risk management and works best combined with other signals.

Category 4: Growth

Revenue growth and earnings growth tend to predict continued outperformance, particularly in growth-oriented market regimes. The market rewards companies that are accelerating, not just companies that are cheap.

Key features: revenue growth (YoY, QoQ), EPS growth, sales acceleration, gross margin trend.

Category 5: Technical and Market Signals

Beyond price momentum, certain shorter-term signals have demonstrated predictive value — particularly relative volume, volatility regimes, and price structure.

Key features: volume relative to average, ATR (average true range), short interest ratio, price relative to sector.

Why Combining Categories Outperforms Any Single Factor

The most important insight from institutional quant research is that multi-factor models consistently outperform single-factor models. Value and momentum are often negatively correlated — when value is working, momentum sometimes isn't, and vice versa. Combining them produces a smoother, more consistent return profile across different market environments.

Machine learning is particularly well-suited to multi-factor modeling because it can discover non-linear interactions between features — combinations that don't show up in simple linear regressions but have real predictive power.

The Data Quality Constraint

None of this works if your data has look-ahead bias. Fundamental data — especially earnings, revenue, and analyst estimates — gets revised after the fact. If your model trains on revised data, it's learning from information that didn't exist at trade time. Point-in-time data, which records exactly what was known on each date, is the prerequisite for trustworthy multi-factor modeling.

How Quant-Builder.ai Puts This Into Practice

Quant-Builder provides 600+ pre-built, point-in-time features spanning all five categories above — valuation, earnings quality, momentum, growth, and technicals — across 3,000+ stocks and 30 years of history. The model-building interface lets you select which feature categories to include, and the platform's feature importance output shows you which features your specific model is actually relying on.

You don't need a quant PhD or a data engineering team. You need a clear thesis — "I want to build a value + momentum model" — and the platform handles the rest.

Quant-Builder.ai — Simplifying Quant Trading: Try a Free Demo at quant-builder.ai/learn

Start Building with the Right Features

The free demo at Quant-Builder.ai lets you explore the full feature library and build a model using any combination of fundamentals, technicals, and macro data — before spending anything. Plans start at $25/month when you're ready to trade it live.