Dispersion Across Trend Programs (Speed/Universe) And Anti-Recency Process Discipline
Sources: 1 • Confidence: Medium • Updated: 2026-02-09 16:40
Key takeaways
- Nick Baltas argued that changing quantitative strategy design reactively in response to recent over- or under-performance is a poor approach and that discipline and process trust are essential.
- Niels Kaastrup-Larsen stated he may raise this paper in a future conversation with Katie to contextualize Crisis Alpha.
- Diversifying trend with non-trend signals (mean-reversion, macro indicators, carry) can help investors maintain allocations during inevitable trend underperformance episodes.
- Investor narratives and thematics may be increasingly influential in shaping asset-allocation decisions and investor views.
- Nick Baltas stated that the incremental value of neural-network-driven non-linear sizing versus a simple pre-specified sigmoid is uncertain and should be evaluated against added complexity and estimation error.
Sections
Dispersion Across Trend Programs (Speed/Universe) And Anti-Recency Process Discipline
- Nick Baltas argued that changing quantitative strategy design reactively in response to recent over- or under-performance is a poor approach and that discipline and process trust are essential.
- Niels Kaastrup-Larsen stated that investors should be cautious about concluding small-universe trend/replicator approaches are structurally superior because in an earlier era (post-tech-bubble to 2009) small universes were consistently the worst performers.
- Nick Baltas stated he suspects 2025 was among the largest dispersion years for trend outcomes, despite generally elevated correlations across trend speeds.
- Dispersion across trend programs in 2025 was driven primarily by differences in trading speed and universe breadth rather than position-sizing choices.
- A 25-year ranking exercise suggests the last three years were uniquely favorable to slow trend signals (roughly 9–12 month speeds), a pattern not seen elsewhere in the sample.
- In the analysis discussed, 2022 ranked very fast trend speeds as best, even though absolute returns were strong across many speeds due to sustained trends that year.
Crisis Alpha And Theoretical Conditions Beyond Return Autocorrelation
- Niels Kaastrup-Larsen stated he may raise this paper in a future conversation with Katie to contextualize Crisis Alpha.
- Independent trial outcomes can exhibit long streaks without violating independence, so observed runs should not be automatically interpreted as dependence.
- Serial correlation is important but not strictly necessary for trend-following profitability because an IID process can still support positive performance under certain conditions.
- Theoretical analysis discussed decomposes trend-following payoff into a directional component and a timing component reflecting how quickly signals align with price dynamics.
- Theoretical results discussed suggest that even when returns are IID, conditioning on sufficiently large realized drawdowns can create conditions under which trend following has positive expected payoff and can exhibit crisis-alpha-like performance.
- Conditional dependence in volatility alone may be sufficient for trend-following to benefit by enabling predictability of directionality.
Beyond-Trend Dynamics (V-Shapes/Mean Reversion) And Portfolio Adaptation
- Diversifying trend with non-trend signals (mean-reversion, macro indicators, carry) can help investors maintain allocations during inevitable trend underperformance episodes.
- Trend researchers should examine beyond-trend dynamics such as reversion effects instead of rewriting core trend philosophy after difficult periods.
- Fixed income, despite being a strong contributor for trend in 2022, has been among the worst sectors to trade since 2022, illustrating rapid rotation in where trend works best.
- Across roughly the last 20 years, all major futures sectors have been profitable for trend at some point, but the strongest opportunities shift materially over time.
- Nick Baltas argued that recent dispersion is better explained by V-shaped reversals and mean reversion than by an environment of uniformly longer-lasting trends favoring slower systems.
- Nick Baltas stated that dynamic allocation can tilt exposures toward or away from sectors based on intertemporal trend signals when the opportunity set changes.
Narratives As Signals And Diagnostics (Regime-Conditional)
- Investor narratives and thematics may be increasingly influential in shaping asset-allocation decisions and investor views.
- Non-price information such as recycled themes and narratives might help quantify market signal-to-noise beyond return-over-volatility heuristics.
- A plausible narrative (example given: higher Japanese rates implying a stronger yen) can fail because markets may show little or no price sensitivity to that narrative.
- Narrative intensity alone does not imply a significant price response; any narrative-to-price linkage may be conditional on thresholds or regimes (example: inflation above a level).
- Nick Baltas is working on a partnership with a vendor/aggregator of narrative data to explore its use for signaling and strategy design.
- If narrative/news data is available before price formation, it might provide earlier signals than price-based momentum indicators.
Signal-To-Position Transfer Functions And Ml Vs Simple Nonlinearity
- Nick Baltas stated that the incremental value of neural-network-driven non-linear sizing versus a simple pre-specified sigmoid is uncertain and should be evaluated against added complexity and estimation error.
- A referenced non-linear time-series momentum approach argues that a neural network can learn market-specific non-linear transfer functions from past returns to expected future returns while preserving the sign-consistency of trend.
- Trend signals can be mapped from past returns to positions using binary, linear scaling, or sigmoid scaling that flattens exposure in the tails to reflect noise and potential reversion.
- Empirical results discussed suggest moving from binary signals toward linear or sigmoid scaling can improve performance partly by reducing costly flipping when signals hover near zero.
- If a learned transfer function crosses zero at extreme signals, it implies mean-reversion dynamics that may be cleaner to implement as a separate engine rather than embedding inside a trend label.
Watchlist
- Investor narratives and thematics may be increasingly influential in shaping asset-allocation decisions and investor views.
- Non-price information such as recycled themes and narratives might help quantify market signal-to-noise beyond return-over-volatility heuristics.
- Trend researchers should examine beyond-trend dynamics such as reversion effects instead of rewriting core trend philosophy after difficult periods.
- Diversifying trend with non-trend signals (mean-reversion, macro indicators, carry) can help investors maintain allocations during inevitable trend underperformance episodes.
- Nick Baltas stated that the incremental value of neural-network-driven non-linear sizing versus a simple pre-specified sigmoid is uncertain and should be evaluated against added complexity and estimation error.
- Niels Kaastrup-Larsen stated he may raise this paper in a future conversation with Katie to contextualize Crisis Alpha.
Unknowns
- What specific narrative datasets, feature definitions (intensity/dispersion/sentiment/topic), and timestamps are used in the referenced narrative-data partnership work, and how are they aligned to tradable horizons?
- What empirical evidence (out-of-sample results, stability across assets, and cost-adjusted performance) supports narrative-based signals improving upon price-based trend signals?
- What is the precise operational definition of 'strong trend conditions' used for the January 2026 characterization, and how is it measured across markets and speeds?
- How large was cross-manager dispersion in 2025 in economically meaningful terms (e.g., top-minus-bottom return spreads), and how does it compare to prior years under consistent universes and cost assumptions?
- What exact speed buckets and universe definitions were used in the 25-year ranking exercise, and how sensitive are results to market set changes, rebalancing rules, and transaction-cost modeling?