Every trading model makes implicit claims about the structure of price data. It claims that certain patterns are signals worth responding to, that others are noise worth ignoring, and that the rules built to exploit the former will continue to work when applied to data the model has never seen. Overfitting and underfitting are the two ways those claims can be wrong, and both have consequences that are difficult to detect until the damage is done. For trend followers targeting Outliers, the challenge has an additional dimension: the very events the process is designed to capture are irregular, non-repeating, and structurally resistant to the optimisation techniques that work well in other contexts.
Signal, Noise, and the Nature of the Edge
Before examining overfitting and underfitting directly, it is worth being precise about what a trading model is actually trying to do. Price data contains two components: signal and noise. The signal is the enduring, exploitable pattern that the model is designed to detect and respond to. The noise is everything else: the random fluctuations, the one-off idiosyncrasies, and the transient patterns that existed in the historical record but carry no predictive information about the future.
For a trend follower, the signal is a directionally trending price series, a sustained move in one direction that reflects a genuine shift in the balance of buying and selling pressure across a meaningful period. Convergent price behaviour, the oscillation of prices around a central equilibrium, is noise from the perspective of a trend-following system. It is not noise in an absolute sense. A mean-reversion strategy treats convergent price behaviour as its signal and directional trends as its noise. The relationship between signal and noise is not a property of the data alone. It is a property of the data relative to the system being deployed.
The signal-to-noise ratio (SNR) of a price series describes how prevalent the system’s target signal is relative to the background noise. A high SNR means that the trending signals the system is designed to exploit are frequent and clear relative to the noise. A low SNR means the noise dominates, and applying the system to that price series produces results driven more by luck than by edge. A system that appears to generate consistent profits from a low-SNR price series is a system that is responding to noise rather than signal. This is not a performance attribute. It is a warning sign.
The edge of a trend-following system does not reside in the system itself. It resides in the price data: specifically, in the enduring and repeatable trending signals that the system is designed to extract. A system operating on a price series with no exploitable trending signal will not generate long-term profits regardless of how well it is constructed. Expecting otherwise is the equivalent of expecting a perpetual motion machine to generate energy from nothing. Understanding this principle is the foundation for understanding overfitting, underfitting, and optimal fitting.
The Tom Basso Experiment: What Random Entry Reveals
A widely cited experiment in trend following methodology challenges some conventional assumptions about the importance of entry signals. Tom Basso, and separately David Harding, demonstrated that a trend-following model with random entry points could still generate long-term profitability. This result initially appears to undermine the importance of entry signal design, but examined closely it reveals something more precise.
Basso’s experiment was not purely random. The entry points were random, but the exit rules were not. The model applied an initial stop and a trailing stop, both of which were specifically designed rather than randomly generated. These exit rules introduced an asymmetry into the system: losses were bounded by the initial stop, and profits were allowed to run until the trailing stop was triggered. This asymmetry, cutting losses short and letting profits run, is not a neutral property. It is a structural alignment with the properties of trending price series. Applied to a price series with genuine trending characteristics, it extracts a profitable edge regardless of where in the trend the entry occurs, because the exit rules ensure that the system participates in the trending move and exits when it reverses.
The implication is precise: the edge in a trend-following system is carried primarily by the exit rules and their alignment with the signal in the price data, not by the precision of the entry. This does not make entry design irrelevant. It means that the asymmetric exit structure is the mechanism through which the edge is actually realised, and that the price series must contain genuine trending signals for that mechanism to extract a positive return.
Three States of Fit
With the signal/noise distinction and the location of the edge established, the three states of model fit can be defined precisely.
An optimally fit model responds to the signals present in the price data while filtering out the noise. Applied to future data, it exploits the majority of the enduring trending signals the price series contains. It does not react to every price pattern that produced a profitable outcome in the historical record. It reacts to the subset of patterns that reflect genuine signal rather than historical noise. This is the desired state.
An overfit model extracts both signal and noise from the historical data. It has been calibrated to respond to all price patterns that produced profitable outcomes in the backtest, including the patterns that were products of random fluctuation rather than enduring signal. The problem with an overfit model is that the noise patterns it has been calibrated to exploit will not recur in the future in the same form. The historical price data contains noise that is specific to that history. Future data will contain different noise. The overfit model, calibrated to respond to historical noise, will fail on future data not because the market has changed in some fundamental way but because the specific random patterns it was trained on no longer exist.
An underfit model fails to fully extract the signal from the price series, leaving a proportion of the exploitable edge uncaptured. Unlike an overfit model, an underfit model will not necessarily deteriorate when applied to future data, because it has not been calibrated to historical noise. Its weakness is insufficiency rather than miscalibration. For trend following specifically, where Outliers are unique, non-repeating events that by definition have not appeared in the historical record in the precise form they will take in the future, a degree of underfitting is not merely acceptable. It is structurally appropriate. An optimally fit trend-following model is one that is deliberately underfit relative to the historical data, with loose enough parameters to accommodate Outlier forms that have not previously occurred.
The Goldilocks framing makes this intuitive. The overfit model is too hot: exquisitely calibrated to a historical record that will not repeat. The underfit model is too cold: insufficiently calibrated to capture the edge that is genuinely present. The optimally fit model is just right: calibrated to the signal while deliberately loose enough to accommodate the range of forms that signal might take in the future. For trend followers targeting Outliers, the just-right state sits closer to the underfit end of the spectrum than it does for strategies targeting more predictable, recurring signals.
Sampling Bias and the Selection Problem
Overfitting and underfitting are compounded by a related problem that arises whenever a choice must be made between alternative models: sampling bias. Consider two traders, Sam and Joe, each developing trend-following systems from the same thirty-year historical dataset, with the most recent five years reserved for out-of-sample testing. After development, both apply their models to the reserved five years. Sam’s model outperforms Joe’s. The natural conclusion is that Sam’s model is superior, and the rational response is to trade Sam’s model going forward.
The problem is that Sam’s model’s superior out-of-sample performance may not reflect superior design. It may reflect the fact that the specific noise patterns in the reserved data happened to align more closely with Sam’s model than with Joe’s. Both models contain a mixture of genuine edge and noise-driven performance, but the selection process cannot distinguish between them. By choosing Sam’s model and discarding Joe’s, a bias is introduced: the selected model is the one whose noise happened to align with the test period, not necessarily the one with the stronger underlying signal extraction capability.
This is the sampling bias problem. It cannot be fully eliminated whenever a choice is made between alternatives, because the choice criterion will always be influenced by the noise component of performance as well as the signal component. The practical response is not to attempt to eliminate sampling bias entirely, but to reduce the role of luck in the selection process through the methods described in the following sections.
Five Methods for Reducing Overfitting in Trend-Following Models
The standard techniques for reducing overfitting in quantitative models, including Monte Carlo simulation and traditional out-of-sample testing, are not well-suited to trend-following systems targeting Outliers. Monte Carlo techniques rely on consistent signal frequencies and amplitudes to generate meaningful synthetic data distributions. Trending price series do not have consistent frequencies or amplitudes. Outliers are irregular, non-repeating, and structurally different from the mean of the distribution. Applying Monte Carlo techniques to a trend-following model produces synthetic data that does not reflect the actual distribution of the signal the model is designed to capture.
The five methods appropriate for trend-following models are as follows.
Design-First Logic
The foundational principle of robust trend-following system development is to establish the design logic before optimising any parameters. This means identifying the edge to be extracted, the enduring trending signal in the price data, and encoding the rules that are structurally capable of capturing that edge before any data mining process begins. The Golden Rules of trend following, cutting losses short and letting profits run, applying consistent position sizing across all return streams, and pursuing low correlation among portfolio components, are the design constraints within which parameter optimisation then operates.
Data mining processes that are given performance objectives but no predefined design logic will fit the model to whatever combination of patterns in the historical data produces the desired performance metrics. The result is a model whose design logic cannot be articulated, whose performance is driven by historical noise, and whose future behaviour cannot be predicted from its structure. By establishing the design logic first and using data mining only to optimise the adjustable parameters within that logic, the model’s behaviour on unseen data is constrained by the design rather than by the historical noise it was trained on.
Simple Models with Few Parameters
Complexity is the primary enabler of overfitting. A model with many parameters has many degrees of freedom with which to fit the historical data. The more parameters it has, the more precisely it can be calibrated to the specific noise patterns in the historical record, and the more severely it will fail when those noise patterns do not recur in the future.
Simple models with few parameters have limited degrees of freedom. They can capture the broad structure of the signal, the general tendency for prices to trend directionally over meaningful periods, without being able to resolve the fine-grained noise that will not repeat. This is the loose pants principle: a model with loose parameters is deliberately underfit to any single historical trend form, which means it is structurally positioned to accommodate the range of trend forms that the future may produce, including Outliers that have never appeared in the historical record. Precision is the enemy of robustness in Outlier-targeting systems. Simplicity and flexibility are the design virtues.
Visual Mapping
After applying design-first logic and building simple models, the evaluation of trade outcomes should not rely exclusively on quantitative performance metrics. Performance metrics are aggregates: they summarise outcomes across the full historical record without revealing whether strong performance periods correspond to periods where genuine trending signals were present or to periods where the model was responding to noise.
Visual mapping addresses this by linking the model’s trade activity directly to the characteristics of the price data. If a model generates its strongest results during the historical periods where clear directional trends are visible in the price data, and exhibits lower activity or flat performance during periods where the price data is noisy and non-directional, this alignment provides evidence that the model is extracting signal rather than noise. Conversely, a model that generates strong performance during visually non-trending periods is a model whose performance is driven by noise, regardless of what the aggregate performance metrics show.
Visual mapping also reduces selection bias by establishing a performance criterion that is grounded in the structural properties of the data rather than in the ranking of models by historical return metrics alone. It is not a purely quantitative method. It is a discipline that connects model design logic to market behaviour in a way that quantitative metrics alone cannot achieve.
Multi-Market Evaluation and Increased Sample Size
The most significant practical challenge in developing robust trend-following models for Outlier capture is the small sample size of Outlier events within any single market and timeframe. A thirty-year historical record of a single liquid market may contain a limited number of genuine Outlier events. Building and validating a model on that sample size alone produces confidence intervals too wide to support robust conclusions about whether strong performance reflects genuine edge or luck.
The solution is to test models universally across a wide range of markets and timeframes simultaneously. Outliers are a universal feature of liquid financial markets. They arise in commodities, currencies, bonds, and equities across multiple geographies and regulatory environments. A trend-following model whose design is grounded in the structural properties of trending price series, rather than in the specific historical patterns of any single market, can be applied consistently across the full range of available markets. Doing so transforms the sample size from the limited Outlier history of a single market into the aggregate Outlier history of hundreds of markets across multiple decades, producing a sample large enough to support robust statistical inference.
This universal testing approach requires that models be deliberately underfit to any single market’s historical characteristics. Market data is normalised using ATR-based methods, treating each market’s volatility history as just another instance of a potential future market condition rather than as a specific template to be fitted. The ATR normalisation ensures that position sizing and signal sensitivity are calibrated relative to each market’s current volatility rather than to an absolute price level, making the model genuinely universal in its application.
Utilising the Full Dataset
In many quantitative approaches, out-of-sample testing is the standard validation method: reserve a portion of the historical data from the development process and test the final model on the reserved data to assess whether its performance generalises beyond the training period. For trend-following models targeting Outliers, this approach is not recommended, and the reason is straightforward.
Outlier events are the signal the model is designed to capture. They are infrequent by definition. Reserving a portion of the historical data for out-of-sample testing removes some Outlier events from the development process entirely. Every Outlier excluded from the development data is a valuable instance of the signal that the model will need to navigate in the future. The five methods described above are sufficient to control overfitting without sacrificing any portion of the available data. Using the full dataset provides the maximum possible sample of Outlier events for model development, and the design-first logic, simplicity constraints, visual mapping discipline, and multi-market evaluation provide the overfitting controls that out-of-sample reservation is conventionally used to provide.
Applying the Framework
The methods described above are not independent techniques. They form a coherent framework for developing trend-following models that are robustly calibrated to the signal in price data rather than to the historical noise that surrounds it.
Design-first logic ensures the model’s structure reflects the mechanism of the edge rather than the surface patterns of the historical data. Simple models with few parameters ensure that the model cannot resolve historical noise with sufficient precision to be miscalibrated to it. Visual mapping provides a structural connection between model performance and the characteristics of the price data that quantitative metrics alone cannot establish. Multi-market evaluation expands the sample of Outlier events to the point where the signal can be distinguished from luck with reasonable confidence. And utilising the full dataset ensures that no Outlier event, the most valuable data points in the historical record, is excluded from the development process.
The objective throughout is not to find the model that performed best in the historical record. It is to find the model that is most structurally aligned with the properties of trending price series across the full range of market conditions, including the conditions that have not yet occurred. A model with loose pants, underfit to any single historical pattern but robustly calibrated to the enduring structural properties of trending markets, is the model best positioned to capture the Outliers that will define long-run geometric compounding in the future.