R&D, Past Returns & Stock Returns

Maletic 2019 — when does R&D spending predict outperformance? Back to index
Author Matjaz Maletic
Affiliation Tilburg / Bank of Slovenia
Sample 1975 – 2014
Observations 414,964 firm-months
Method Fama–MacBeth
The single insight

The same R&D number means opposite things depending on whether the stock has just won or lost.

Most people read R&D spending as one signal. Maletic shows it's two: the level of R&D and the change in R&D interact with past returns in opposite directions. High R&D pays off after a down year; rising R&D pays off after an up year. The combination of past performance and R&D behavior is what carries the signal.

Section 1 · The question

Does R&D spending predict future stock returns?

Earlier work (Chan, Lakonishok & Sougiannis 2001) showed that firms with a high R&D-to-market-value ratio (RDM) earn higher future returns. But RDM has a problem: its denominator is market value, which itself reflects past returns. So you can't tell whether RDM is picking up "R&D intensity" or just "stocks that have fallen recently."

Maletic untangles the two. He asks two cleaner questions:

The answers turn out to be very different — and that's the point of the paper.

Section 2 · The result in one picture

Two stories, opposite signs

Story 01 · R&D Level

High R&D + bad year → outperform

high RDA × low past return + future return

Why: A firm that has just performed badly but refuses to cut R&D is signaling management confidence the market is missing. The market over-reacts to the bad year, mispricing kicks in, and the stock catches up later.

Story 02 · R&D Change

Rising R&D + good year → outperform

high ΔR&D × high past return + future return

Why: Increasing R&D is only a credible bullish signal when management has already proven they can pick winners. A bad-performing firm that suddenly ramps R&D is not rewarded — only firms with a positive track record are.

Same R&D, opposite interpretations. The cross-product with past return flips the sign.

Section 3 · The four R&D measures

What each ratio is actually capturing

RDA

R&D-to-Assets

R&D / Total Assets

A clean level measure. Denominator (assets) doesn't move with the stock price, so this isolates R&D intensity from past performance.

RDM

R&D-to-Market Value

R&D / Market Cap

Mixes R&D intensity with past returns (market cap reflects price). Strong predictor — but the source of its power is unclear.

RDGP

R&D-to-Gross-Profits

R&D / (Revenue − COGS)

R&D intensity scaled by raw profitability. In Maletic's regressions, this one has no predictive power on its own.

ΔR&D

Yearly R&D Growth

(R&D_t − R&D_t−12) / R&D_t−12

A change measure. Captures whether the firm is ramping or cutting R&D — a behavioral signal, not a level.

Note the decomposition: Gross Profitability (GPA) = RDA / RDGP. The control variable GPA already contains the level measure, so the regressions cleanly separate "raw profitability" from "R&D intensity."

Section 4 · The method

Fama–MacBeth cross-sectional regressions

STEP 01
Build the sample
Compustat + CRSP, 1975–2014. Drop financials, tiny firms, negative book equity, zero-R&D firms.
STEP 02
Regress monthly
Each month, regress next-month return on RDA, ΔR&D, RDM, GPA, size, B/M, prior-month return.
STEP 03
Add interactions
Add RDA × past_1y_return and ΔR&D × past_1y_return as predictors.
STEP 04
Average the slopes
Average the 451 monthly cross-sectional slopes; Newey–West standard errors (lag 12) for t-stats.

All independent variables winsorized at 1% / 99% to suppress extreme values.

Section 5 · The key numbers

Slope coefficients from Table 3

Predictor Slope t-stat Sig? What it means
RDA × r−2,−12 −0.0499 −2.37 YES High R&D level pays off more after bad past returns. The minus sign confirms the interaction.
ΔR&D × r−2,−12 +0.0094 +4.42 YES Rising R&D pays off more after good past returns. Opposite sign to the level interaction.
ΔR&D (alone) −0.0024 −2.61 YES Without conditioning on past returns, increasing R&D hurts on average — only winners are rewarded for ramping.
RDM +0.0021 +2.49 YES R&D-to-market-value still predicts higher returns even after the interactions are added. Anomaly survives.
RDA (alone) ≈ insignif. +1.20 NO Once you control for RDM, the standalone level loses its bite — most action is in the interaction term.

Both interaction signs are robust to weighted least squares (weighting by firm size). Coefficients shift only marginally.

Section 6 · What the paper failed to show

The R&D-to-market-value anomaly survives

A natural hypothesis (proposed by Chan et al. 2001) is that the RDM premium is really just the "high R&D + bad past returns" story repackaged — because past returns sit in the denominator of RDM. If that were true, adding the RDA × past-return interaction should soak up RDM's predictive power.

Maletic's test: add both interactions to the regression, then check whether RDM's coefficient collapses to zero.

It doesn't. RDM's slope stays at +0.0021 with t-stat +2.50, even with everything else in the model. Whatever drives the R&D-to-market-value premium is not the past-return interaction — it's something else, still unexplained.

Section 7 · Bottom line

Three things to remember

01

Past returns flip the meaning of R&D

Don't treat R&D-to-assets as a single number. The combination with past 1-year return is the actual signal — and the combination has opposite sign for levels vs. changes.

02

The market is sluggish at updating views on past losers

Firms that perform poorly but refuse to cut R&D are systematically under-priced. Their continued spending is a credible bullish signal the market discounts too heavily.

03

RDM still hides a separate anomaly

After controlling for the level/change interactions with past returns, R&D-to-market-value still predicts higher future returns. There is at least one more story in there waiting to be explained.

Reference

Source paper

Maletic, M. (2019). R&D Investments, Past Returns, and the Cross-Section of Stock Returns.
Working paper, Tilburg University Finance Department / Bank of Slovenia Research Department.
SSRN ID 3178186 — version 26 September 2019.

Paper is in the working-paper stage in the version I read; numbers above are from Maletic's Table 3, which uses Newey–West standard errors with lag 12 on Fama–MacBeth (1973) cross-sectional regressions.