Breaking the Trend: How to Avoid Cherry-Picked Signals
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Breaking the Trend: How to Avoid Cherry-Picked Signals".
Warren: Here's the gist of the paper. Our empirical results show an impressive fit with the theoretical Sharpe formula of a trend-following strategy depending on the parameter of the signal, which was derived by Grebenkov and Serror (2014).
Ruth: This empirical fit convinces us that a mean-reversion process with only one time scale is enough to model, in a precise way, the reality of the trend-following mechanism at the average scale of CTAs and as a consequence, using only one simple EMA, appears optimal to capture the trend.
Warren: As a consequence, using a complex basket of different complex indicators as signal, does not seem to be so rational or optimal and exposes to the risk of cherry-picking.
Ruth: Theoretical Framework Validation The main objective of this paper is to empirically validate the theoretical Sharpe ratio model proposed by Grebenkov and Serror (2014) and therefore their assumption to model trends as gaussian mean reversion stochastic processes with only one time scale.
Paper summary: Ruth: So to wrap up this paper Breaking the Trend: How to Avoid Cherry-Picked Signals, the authors really argue that using a complex mixture of sophisticated indicators is unnecessary when the simple Exponential Moving Average provides a perfect fit Ruth.
Warren: They found that it's better than just picking any indicator because it replicates the theoretical Sharpe ratio derived by Grebenkov and Serror two thousand fourteen quite closely Warren <ref:2504.10914#pg1,derived by Grebenkov and Serror 2014>.
Iris: The core implication is that modeling trends with only one time scale and mean reversion isn't just a nice idea, it appears to be a good solution because the empirical fit is more than correct Iris.
Ruth: It means for strategy selection, you should prioritize simplicity when you find that the single EMA indicator gives you such a strong fit Ruth.
Warren: Because if you use that simple EMA with its optimal parameter of one hundred twelve plus or minus ten business days, the results are quite compelling Warren <ref:2504.10914#pg1>.
Iris: It’s surprising how optimal a single EMA turns out to be for capturing trends, especially when you expect different time scales for different investors Iris.
Ruth: So the paper Breaking the Trend: How to Avoid Cherry-Picked Signals basically suggests that simplicity is actually beautiful when it comes to finding an optimal trend signal Ruth.
Conclusion: Ruth: So we've seen how this paper breaks down the math to show that simple signals are often the best ones for capturing trends Ruth.
Warren: It really shows how they took a complex theoretical model, that Sharpe ratio formula from Grebenkov and Serror, and turned it into something you can actually test with real market data Warren.
Iris: The main point there is that relying on a single time scale for mean reversion might be enough to model the whole trend mechanism accurately Iris.
Ruth: It suggests that using a big basket of different complex indicators isn't rational, because the simple exponential moving average fits so well Ruth.
Warren: And when you look at their results, they found that this single EMA signal is optimal with a specific parameter around one hundred and twelve business days Warren.
Iris: That’s interesting because I always think about how different investor types might need different time scales, so finding one best fit here is something to watch Iris.
Ruth: It really gets you thinking about what this means for actual trading decisions when you're trying to pick a strategy Ruth.
Warren: Because if the math works this well, it opens up a new way of thinking about what makes a good signal in these markets Warren.
Machina Capital · Valeyre Research
q-fin.PM
Submitted: 2025-04-15
Updated: 2026-11-18
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 74/100
The gist: Our empirical results show an impressive fit with the theoretical Sharpe formula of a trend-following strategy depending on the parameter of the signal…
Key concepts
- Sharpe Formula Fit
- This refers to how well the empirical performance of a trading strategy matches a theoretical mathematical model derived by Grebenkov and Serror. The paper found that the model fits best when using a single EMA, validating the underlying theory.
- One Time Scale Model
- The theory suggests that trends can be accurately modeled using only one time scale for mean-reversion processes. Empirical testing confirmed this simplicity is sufficient to precisely capture the trend mechanism at average scales.
- Cherry-Picking Risk
- This is the danger of using a complex basket of different technical indicators as signals. The paper argues that relying on many indicators exposes the strategy to cherry-picking, where one indicator appears superior in hindsight, leading to suboptimal performance.
- Exponential Moving Average (EMA)
- The EMA is a simple trend indicator used in the study. The optimal parameter for this EMA was found to be 112 business days, which corresponds to a half-life period of 78 business days for achieving the best Sharpe ratio.
Terminology
Summary
Our empirical results show an impressive fit with the theoretical Sharpe formula of a trend-following strategy depending on the parameter of the signal, which was derived by Grebenkov and Serror (2014). p. 0 This empirical fit convinces us that a mean-reversion process with only one time scale is enough to model, in a precise way, the reality of the trend-following mechanism at the average scale of CTAs and as a consequence, using only one simple EMA, appears optimal to capture the trend. p. 0 As a consequence, using a complex basket of different complex indicators as signal, does not seem to be so rational or optimal and exposes to the risk of cherry-picking. p. 0
Theoretical Framework Validation
The main objective of this paper is to empirically validate the theoretical Sharpe ratio model proposed by Grebenkov and Serror (2014) and therefore their assumption to model trends as gaussian mean reversion stochastic processes with only one time scale. The theoretical Sharpe ratio is derived from a classical one-dimensional model assuming returns consist of two components: a random, independent and identically distributed (i. p. 9i.d.) component, and a small bias term modeled using an auto-regressive process of order 1, AR(1). p. 9 This classical one-dimensional model can be extended to multiple dimensions using two covariance matrices to account for correlations. p. 9 The returns are described by the formula ri,t = εi,t + βPt−1k=1 (1 − λ)t−1−k ξi,k. p. 9
Empirical Testing Methodology
The authors conducted empirical backtests of the Agnostic Risk Parity (ARP) strategy introduced by Benichou et al (2017) to validate Grebenkov’s theory in Eq. (3). They focused primarily on exponential moving averages (EMA) as trend indicators, testing various EMA configurations and more complex signals like the Moving Average Convergence Divergence (MACD) to confirm that one time scale is enough to describe trends. p. 6 The simulation started on May 29, 1990, and stopped on December 7, 2023, using daily returns from 70 futures instruments in stock indices, bonds, FX and commodities futures. p. 7
Performance of the Optimal Signal
The empirical simulation for ARP fits pretty well the theorical formula of the Sharpe ratio Eq. (3) derived in Grebenkov and Serror (2014) with parameters λ = 1/180±17 and β0 = 0. p. 012. The parameter of 112 ± 10 business days (equivalent to a half-life period of 78 business days) for simple EMA is the optimal parameter to get the optimal Sharpe ratio when not accounting cost of trading which are very small at that trading frequency. p. 4 The empirical Sharpe ratio for ARP (80) EMA is 1. p. 6240349, and for MACD (20, 80, 0. p. 33 × 400) it is 1. p. 8218031.
Robustness Against Signal Complexity
The paper demonstrates that a simple Exponential Moving Average (EMA) signal can be replicated through a complex mixture of Bollinger Bands (BB) indicators. Furthermore, the empirical Sharpe ratio closely replicates the theoretical Sharpe ratio predicted by Grebenkov and Serror (2014), showing that a single EMA indicator is already optimal. p. 6 The analysis also shows that the ARP approach yields higher Sharpe ratios than the naive ’1/N’ approach, even when binary trading rules are used. p. 5
Implications for Strategy Selection
The conclusion is that using a complex mixture of sophisticated indicators is unnecessary when the EMA alone provides a perfect fit, proving that simplicity can indeed be beautiful. The findings suggest that modeling trend through mean-reversion process using only one relaxation time and not an multi-time scales one appears to be a good solution as the fit is more than correct. p. 0 This suggests that using a complex basket of different complex indicators as signal, does not seem to be so rational or optimal. p. 0
Conclusion
Grebenkov’s model for describing trends is empirically validated, as its rather complicated theoretical formula for determining the Sharpe ratio based on the EMA parameter fits impressively well with empirical data. The best fit is obtained using the theoretical model parameters λ = 1/180±17 and β0 = 0. p. 112. As a consequence, the parameter of 112 ± 10 business days (equivalent to a half-life period of 78 business days) for simple EMA is the optimal parameter to get the optimal Sharpe ratio. p. 4 It is quite surprising that a single EMA is optimal for capturing trends, as one would expect different time scales for different types of investors or long-term mean reversion at longer time scales. p. 7 The conclusion is that using a complex mixture of sophisticated indicators is unnecessary when the EMA alone provides a perfect fit—proving that simplicity can indeed be beautiful. p. 0 p. 0
References
The paper cites numerous works to support its theoretical and empirical claims, including Grebenkov and Serror (2014), Benichou et al (2017), and various literature on mean-variance optimization uncertainty such as Ledoit and Wolf (2003, 2004). The paper also references specific technical indicators like the SMA, EMA, and MACD, and studies on regime-switching models such as Zakamulin and Giner (2023). p. 4 p. 4 The empirical results are validated by backtesting over a long period, comparing the ARP portfolio to the naive ’1/N’ approach. p. 6 p. 6 The paper also discusses non-linear findings from Schmidhuber (2021) and various theoretical formulations of optimal portfolios incorporating estimation uncertainty. The list of cited references is extensive, covering topics from classical portfolio theory to modern regime-switching models in finance. p. 4 p. 4 The paper also details the specific data used for simulation, including commodity futures and stock indices. p. 7 p. 7 p. 7 The final section provides a comparison of Grebenkov’s formula with other literature, such as Acar (2003). p. 6 p. 6 The paper also includes a detailed list of the 32 commodities futures traded in the USA or in Europe. p. 1 p. 1 p. 1
Data and Instruments
Table 1 lists the 32 commodities futures traded in the USA or in Europe. Table 2 lists the 41 instruments among currencies, equity indices and Bonds. p. 7 The simulation starts on May 29, 1990, and stops on December 7, 2023. p. 4 The data used includes daily returns from 70 futures instruments in stock indices, bonds, FX and commodities futures. p. 7
Key Findings on Portfolio Construction
The paper emphasizes the importance of constructing a rotationally invariant portfolio through the Agnostic Risk Parity (ARP) strategy. Implementing the ARP portfolio involves normalizing trend indicators by the inverse square root of the correlation matrix, rather than by the full inverse as in the classical Markowitz framework. p. 2 The optimal strategy is found to be a single EMA indicator when accounting for trading costs. p. 7 The scaling factor formula, which depends on the size of the universe N, empirically validates the theoretical scaling of the Sharpe ratio for a portfolio of N assets. p. 1 The results show that diversification and data aggregation—through projection onto the ARP—help reduce the noise-to-signal ratio of measurements when N is large. p. 6
Theoretical Extensions and Limitations
The theoretical model, while elegant, has limitations; for instance, it may be specific to the Dow Jones Index. The formula in Eq. p. 0 (3) is derived for the one-dimensional case only but it could be easily extended to the multi-dimensional case. p. 5 The fit may be specific to the Dow Jones Index, as the fit is not perfect for time scales shorter than 100 days. p. 0 Furthermore, the model assumes that positions should better depend linearly on the signals when assuming that returns are well described by Eq. p. 5 (1). The results show that market behavior is not necessarily governed by a multi-time scales property in the way some literature suggests. p. 0 The paper notes that the empirical Sharpe ratio is 1. p. 124 for ARP (80) with one time scale and 1. p. 618 for MACD (20, 120, 0 × 400). p. 8 The conclusion is that using a complex mixture of sophisticated indicators is unnecessary when the EMA alone provides a perfect fit—proving that simplicity can indeed be beautiful.
Mathematical Derivations
The derivation of the theoretical Sharpe ratio involves complex mathematical expressions, such as Eq. (3). The asymptotic variance of the returns in Eq. p. 2 (23) is derived by rescaling the parameter β in order to make the asymptotic variance independent of the timescale λ by defining: β0 = βp/λ(2 − λ). p. 2 The closed-form formula for the Sharpe ratio SR(H) in terms of model parameters is given by Eq. p. 1 (37). The paper also explores how a simple Exponential Moving Average (EMA), which is supposed to be both optimal and simple, could also be replicated by a highly complex but more usual combination of indicators, such as Bollinger Bands. p. 7 The analysis includes non-linear findings from Schmidhuber (2021) which challenge the assumption of a purely gaussian diffusive return process. The paper also explores how the optimal signal depends on a nonlinear transformation of the form ϕ − cϕ3, with c = 0. p. 133.
The paper cites numerous works to support its theoretical and empirical claims, including Grebenkov and Serror (2014), Benichou et al (2017), and various literature on mean-variance optimization uncertainty such as Ledoit and Wolf (2003, 2004). The paper also references specific technical indicators like the SMA, EMA, and MACD, and studies on regime-switching models such as Zakamulin and Giner (2023). The empirical results are validated by backtesting over a long period, comparing the ARP portfolio to the naive ’1/N’ approach. The paper also discusses non-linear findings from Schmidhuber (2021) and various theoretical formulations of optimal portfolios incorporating estimation uncertainty. The list of cited references is extensive, covering topics from classical portfolio theory to modern regime-switching models in finance. The paper also details the specific data used for simulation, including commodity futures and stock indices. The final section provides a comparison of Grebenkov’s formula with other literature, such as Acar (2003). The paper also includes a detailed list of the 32 commodities futures traded in the USA or in Europe. The paper also details the specific data used for simulation, including commodity futures and stock indices. The paper also includes a detailed list of the 32 commodities futures traded in the USA or in Europe.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Implement a theoretical Sharpe ratio calculation using Grebenkov and Serror (2014) formula, which is described as
a remarkably elegant formula,
to derive optimal signal parameters from empirical backtests. This allows the AI system to select theoptimal smoothing parameter
such as 1/112 business days for a simple EMA, rather than relying on purely empirical optimization. -
Develop a portfolio construction module that uses the Agnostic Risk Parity (ARP) strategy, which is described as
theoretically and empirically considered to be the optimal trend-following portfolio.
This module can construct rotationally invariant portfolios by normalizing signals through the inverse of the square root of the correlation matrix, making performance less dependent on subjective selection of futures. -
Create a signal processing layer that tests whether
one time scale is enough to describe trends
by comparing an Exponential Moving Average (EMA) against complex mixtures like Bollinger Bands or MACD, and then selecting the simplest indicator—the EMA—if it provides aperfect fit
as shown in the empirical results. -
Integrate a robust uncertainty modeling component that extends the Grebenkov model by treating the trend component not as a constant but as
a path-dependent quantity that depends on Pt-1k=1(1-λ)t-1-kξi,k,
allowing the system to capture non-Gaussian return processes. -
Implement a scaling factor validation tool that empirically tests the formula suggesting that the Sharpe ratio depends on the size of the universe as
q N 1+(N−1ρ2),
enabling an AI to predict expected Sharpe ratios for different portfolio sizes (N) based on empirical data rather than just theoretical scaling.
Sources
- Black was right: Price is within a factor 2 of Value
- Is Trend Still Your Friend?: A Microstructural Account of the Demise of Short-Term Trend-Following
- Trends and Reversion in Financial Markets on Time Scales from Minutes to Decades
- Basis Immunity: Isotropy as a Regularizer for Uncertainty
Related papers
- Bayesian Parametric Portfolio Policies
- Mean--Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis, and Empirical Study
- Beyond Co-Movement: Locality by Exposures Enables a Joint Factor-Graph Framework for Portfolio Diversification
- Interpretable Allocation and Model-Predictive Rebalancing of Equity Portfolios, Tested on Real and Synthetic Markets
- Penalize the Common Trade: A Collision-Aware Layer for Learned Portfolios
- Mitigating Return Drag and Regime Whipsaws in Sector Rotation: An Asymmetric Downside-Aware Nested Clustered Optimization Framework