Evidence-First Factor Investing
Test investment claims across clean data and combine robust signals
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 98%
Begin with a concrete claim about why certain securities should outperform, then test it using a comprehensive point-in-time database. Include bankrupt and acquired companies, and only expose each observation when an investor could actually have known it. Examine the signal across market cycles and identify environments in which its behavior changes, such as momentum reversing after severe bear markets. Retain factors that demonstrate both empirical persistence and intuitive economic logic. Rather than allowing one metric to determine the portfolio, combine several complementary factors into a composite. Continue cleaning the data and researching the model, but make incremental changes so that short-term noise does not displace durable foundations.
Origin
Jim O'Shaughnessy developed the approach through early Dow research and the work that became What Works on Wall Street, later applying it at Bear Stearns and O'Shaughnessy Asset Management.
Core principles
- 01Demand empirical evidence for every investment claim
- 02Eliminate survivorship and look-ahead bias
- 03Prefer economically intuitive factors
- 04Combine complementary factors instead of trusting one signal
- 05Improve models evolutionarily rather than chasing reinvention
How to run it
- 1
State a testable claim
Translate an investment belief into a measurable rule, such as buying companies with low valuations or strong momentum.
Pro tip Choose claims with an intuitive economic mechanism rather than arbitrary correlations.
Watch out Do not begin with a desired result and search backward for supporting variables.
- 2
Build a point-in-time universe
Include active, acquired, and bankrupt companies so the historical universe reflects what investors actually faced.
Pro tip Preserve delisted securities and historical index membership.
Watch out A current-company database introduces survivorship bias.
- 3
Lag the information
Make each fundamental value available only after its historical publication date.
Pro tip Model conservative reporting delays when exact release dates are unavailable.
Watch out Using period-end figures before publication creates look-ahead bias.
- 4
Test across regimes
Measure performance through expansions, recessions, bubbles, crashes, and recoveries. Look specifically for conditions that invert or weaken the signal.
Pro tip Separate ordinary corrections from severe bear markets.
Watch out An aggregate long-run result can conceal destructive regime dependence.
- 5
Form a composite
Combine several validated factors so the portfolio does not depend on one fragile metric.
Pro tip Favor factors that capture different dimensions such as value, quality, and momentum.
Watch out Adding correlated factors can create the illusion of diversification.
- 6
Clean and evolve
Reconcile questionable records and refine the model as better evidence emerges. Preserve the durable foundation while making measured improvements.
Pro tip Document why every factor or data rule changes.
Watch out Frequent revolutionary changes can turn research into performance chasing.
In the wild
O'Shaughnessy's research found that inexpensive smaller companies became a stronger selection set when their prices had begun improving. The composite joined valuation with evidence that the business or market perception had turned a corner.
→ A more selective small-value strategy than buying every statistically cheap company.
Longer historical research covering the Great Depression showed that many low-price-to-book companies subsequently failed. The team used that evidence to remove price-to-book from its composites rather than preserving a traditional factor by convention.
→ The model was revised after a historically important failure mode became visible.
Common mistakes
Testing only survivors
Restricting research to companies that exist today removes many historical failures and inflates apparent returns.
Using information too early
Assigning a reported number to its fiscal date rather than its public release date gives the model knowledge investors did not possess.
Trusting one factor
A single metric can fail because of structural changes, unusual regimes, or weaknesses specific to that measure.
Is it for you?
Best for
It is best for investors and researchers building systematic equity-selection models.
Not ideal for
It is not ideal for investors without sufficient historical data, statistical discipline, or tolerance for extended underperformance.
From the transcript
“Where is the proof? Where is the empirical evidence”
“We don't use a single factor.”
“What we're always trying to do is improve them. But it's evolutionary, not revolutionary.”
From the episode
Jim O'Shaughnessy on How AI Will Change Everything From Arts to Stocks