How does time-period bias affect CME research, and why are findings so sensitive to start and end dates?
The CFA Level III material mentions that research results can change dramatically depending on which dates you pick. How big is this problem really, and what can analysts do about it?
Time-period bias occurs when research findings are specific to the particular start and end dates chosen for the analysis. The bias arises because different periods capture different market regimes, economic conditions, and structural features — and conclusions drawn from one window may not hold in another.
How Sensitive Are Results?
Extremely. Consider the question of whether value stocks outperform growth stocks:
| Period | Value vs. Growth Annual Spread | Conclusion |
|---|---|---|
| 1940–1975 | +3.8% | Value dominates |
| 1990–1999 | -8.2% | Growth dominates (tech boom) |
| 2000–2006 | +6.1% | Value dominates (tech bust recovery) |
| 2017–2020 | -12.4% | Growth dominates (FAANG era) |
| 2022–2023 | +5.3% | Value dominates (rate hikes) |
An analyst who picks any single window gets a definitive answer — but a completely different one depending on which window they choose.
Example — Oakmont Asset Management:
Oakmont's CIO commissions research on whether momentum strategies work in emerging markets. The research team finds:
- 2005–2015: Momentum earns +7.3% annual alpha (strong result)
- 2015–2020: Momentum earns -1.8% annual alpha (reversal)
- 2005–2020: Momentum earns +3.4% annual alpha (moderate)
The CIO asks: 'Should we implement a momentum tilt in our EM allocation?'
The answer depends entirely on which period the analyst emphasizes. If the 2015–2020 reversal represents a permanent structural change (e.g., faster information diffusion via technology), the earlier alpha may not recur. If 2015–2020 was a temporary anomaly (e.g., driven by a unique commodity cycle), the longer sample is more informative.
Defenses Against Time-Period Bias:
- Test across multiple sub-periods: If a relationship only works in one window, it's fragile
- Use rolling windows: Instead of a single start-to-end calculation, compute rolling 5-year or 10-year estimates and examine stability
- Test across different markets: A relationship that holds in the US, Europe, and Asia is more credible than one limited to a single market
- Combine with economic logic: If there's a sound reason for the relationship AND it holds across multiple periods, it's more likely genuine
- Report sensitivity: Always show how results change with different start/end dates
Key Exam Insight: Time-period bias and data-mining bias often work together. An analyst who cherry-picks the start and end dates that produce the strongest result is effectively data-mining along the time dimension.
For more on CME biases, explore our CFA Level III community Q&A.
Master Level III with our CFA Course
107 lessons · 200+ hours· Expert instruction
Related Questions
What exactly is the Capital Market Expectations (CME) framework and why does it matter for asset allocation?
How do business cycle phases affect asset class return expectations?
Can someone explain the Grinold–Kroner model step by step with numbers?
How do you forecast fixed-income returns using the building-blocks approach?
PPP vs Interest Rate Parity for forecasting exchange rates — when do I use which?
Join the Discussion
Ask questions and get expert answers.