Hypothesis-First Quantitative Investing
Test economic explanations with robust data instead of mining patterns
- Difficulty
- Expert
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 95%
Hypothesis-First Quantitative Investing treats quantitative tools as instruments for economic inquiry rather than pattern-mining machines. Researchers begin with competing hypotheses about a market effect and identify the behavioral, incentive, or risk mechanism each predicts. They then gather relevant data, test those predictions, and subject promising results to robustness checks. Model construction is intentionally intensive up front because the goal is to isolate the simplest durable signal, not continually add variables after observing new noise. Gray emphasizes persistent human behavior and principal-agent incentives as more stable foundations than transient macro commentary. Once deployed, the model remains open to disconfirming evidence, but changes require a material failure of the mechanism or assumptions. This discipline avoids the gradual accumulation of small modifications that can transform a coherent strategy into something unrelated to its original thesis.
Origin
Gray contrasts Alpha Architect's economist-led research with stereotypical quants who data-mine complex models without financial understanding. Extracted from Masters in Business.
Core principles
- 01Start with competing explanations for how the world works
- 02Use data to test hypotheses rather than manufacture stories
- 03Prefer robust signals to elaborate models
- 04Human behavior and incentives are persistent explanatory inputs
- 05Guard against model creep
How to run it
- 1
State the mechanism
Explain the human behavior, incentive, risk, or market friction that could generate the proposed return pattern.
Pro tip Write the explanation before viewing the final result.
Watch out A story invented after seeing the data may merely rationalize noise.
- 2
Form competing hypotheses
Specify alternative explanations and the different observations each would predict.
Pro tip Include a null hypothesis under which the apparent edge is random.
Watch out Testing only a favored explanation creates confirmation bias.
- 3
Collect discriminating evidence
Use data that can meaningfully separate the hypotheses rather than simply produce an attractive backtest.
Pro tip Reserve independent periods or markets for out-of-sample evaluation.
Watch out Repeatedly searching the same dataset increases false discoveries.
- 4
Stress-test the result
Vary samples, assumptions, definitions, and implementation details to determine whether the signal survives reasonable perturbations.
Pro tip Look for consistency of direction and mechanism, not one perfect specification.
Watch out A result dependent on one parameter choice is unlikely to be robust.
- 5
Simplify the model
Retain the strongest explanatory signal and remove complexity that primarily fits historical noise.
Pro tip Compare every added feature with a simpler baseline.
Watch out Complexity can disguise overfitting while creating false confidence.
- 6
Control model changes
Monitor whether the underlying mechanism and assumptions remain credible, but require strong evidence before altering the deployed model.
Pro tip Document each proposed change and test it independently.
Watch out Incremental model creep can produce a strategy that no longer represents the original thesis.
In the wild
A researcher hypothesizes that investors overreact to bad news, pushing disliked stocks below fundamental value. The researcher tests cheap stocks across periods and markets, examines whether results persist after risk controls, and compares simple valuation definitions before deploying a restrained model.
→ The resulting strategy has an explicit mechanism and survives broader scrutiny than a correlation discovered by unconstrained searching.
A new variable improves one historical backtest but lacks an economic mechanism and disappears with a modest change in dates. The team rejects it rather than adding it to the production model.
→ Model simplicity and out-of-sample reliability are preserved.
Common mistakes
Mining first and explaining later
Searching unlimited variables can produce impressive accidental patterns that lack a durable mechanism.
Equating complexity with sophistication
More variables can fit more historical noise. The objective is the clearest robust signal, not the most elaborate specification.
Tweaking after every disappointment
Small reactive changes accumulate into model creep and can destroy the original economic logic.
Is it for you?
Best for
Researchers building systematic strategies from financial, economic, or behavioral evidence.
Not ideal for
Traders seeking opaque short-lived correlations without a defensible causal or behavioral explanation.
From the transcript
“So we have competing hypotheses about how the world works. We grab data, we test these competing hypotheses, and we're always trying to get better…”
“Simpilicity beats complexity definitely, because we want to look for the what is the real signal here, not the noise, and let's just focus on…”
“Human behavior is pretty constant, and incentives, specifically like this principal agent delegating asset management problem, those are two constants”
From the episode
Interview With Wesley Gray: Masters in Business (Audio)
Wesley Gray