Navigating Shifting Markets: A Causal Approach to Smarter Trading

Author: Denis Avetisyan

A new strategy leverages causal inference and forward-looking indicators to adapt to the ever-changing dynamics of financial markets and improve portfolio performance.

A normalization process-applied to financial indicators including Money Flow Index, Relative Strength Index, Bollinger Bands percentage, and Moving Average Convergence Divergence for both EURUSDT and BTCUSDT-centers the data by subtracting a 5000-minute rolling median <span class="katex-eq" data-katex-display="false">I~^{(k)}\_{t}=I^{(k)}\_{t}-m^{(k)}\_{t}</span>, then scales it using the rolling Median Absolute Deviation <span class="katex-eq" data-katex-display="false">s^{(k)}\_{t}</span> to produce a dimensionless, comparable metric <span class="katex-eq" data-katex-display="false">Z^{(k)}\_{t}=\tilde{I}^{(k)}\_{t}/s^{(k)}\_{t}</span> that facilitates direct cross-indicator analysis and aggregation. — A normalization process-applied to financial indicators including Money Flow Index, Relative Strength Index, Bollinger Bands percentage, and Moving Average Convergence Divergence for both EURUSDT and BTCUSDT-centers the data by subtracting a 5000-minute rolling median $I~^{(k)}\_{t}=I^{(k)}\_{t}-m^{(k)}\_{t}$ , then scales it using the rolling Median Absolute Deviation $s^{(k)}\_{t}$ to produce a dimensionless, comparable metric $Z^{(k)}\_{t}=\tilde{I}^{(k)}\_{t}/s^{(k)}\_{t}$ that facilitates direct cross-indicator analysis and aggregation.

This review details a performance-driven causal signal engineering framework for financial markets operating under non-stationarity, utilizing walk-forward optimization and causal observables.

Traditional financial modeling often struggles with the inherent non-stationarity of market dynamics, limiting adaptive performance. This paper, ‘Performance-Driven Causal Signal Engineering for Financial Markets under Non-Stationarity’, introduces a novel framework for constructing strictly causal trading signals from heterogeneous indicators, dynamically adapting to regime shifts via walk-forward optimization. The resulting strategy demonstrates a pronounced risk-reshaping effect, achieving smoother equity trajectories and reduced drawdowns under controlled conditions, effectively establishing an upper bound on achievable performance. Could this approach to causal signal engineering unlock new strategies for navigating complex, non-stationary systems beyond financial markets?

Navigating the Inherent Uncertainty of Financial Markets

Financial markets differ fundamentally from many physical systems in that their statistical properties are not constant over time; they are, by nature, non-stationary systems. This means that relationships observed between financial variables – correlations, means, variances – are subject to change, often unpredictably. Traditional statistical models, predicated on the assumption of stationary data, therefore struggle to accurately capture or forecast market behavior. A model calibrated to historical data may quickly become unreliable as market conditions evolve, leading to flawed predictions and potentially significant financial losses. The inherent dynamism of these systems necessitates the development of adaptive modeling techniques and a continuous reassessment of model parameters to account for these shifting statistical landscapes. This characteristic makes consistently profitable trading incredibly challenging, requiring strategies that acknowledge and respond to the ever-changing nature of financial data.

Financial markets aren’t static; they undergo frequent regime transitions – distinct shifts in behavior characterized by altered volatility, correlation, and overall market dynamics. These transitions, often triggered by macroeconomic events, policy changes, or even shifts in investor sentiment, necessitate strategies that move beyond relying on historical data alone. Traditional models, calibrated to past relationships, can quickly become ineffective as markets jump between these regimes – for example, from periods of low volatility and positive returns to those of high volatility and negative returns. Consequently, adaptive strategies – those capable of identifying and responding to these shifts – are paramount. These might include dynamic asset allocation, time-varying parameter models, or machine learning algorithms designed to continuously recalibrate based on incoming data, allowing systems to capture the evolving nature of financial markets and maintain robustness across different economic environments.

Financial market robustness hinges on a deep comprehension of underlying dynamics and the pervasive influence of volatility. These aren’t simply random fluctuations; rather, volatility represents the rate of change in price, driven by information flow, investor sentiment, and macroeconomic factors. A trading system designed without accounting for these shifts risks quickly becoming obsolete as market conditions evolve. Successful systems, therefore, incorporate mechanisms to not only measure volatility – often using metrics like standard deviation or $implied volatility$ – but also to adapt to its changing characteristics. This can involve dynamically adjusting position sizes, employing trend-following strategies, or utilizing options to hedge against unexpected price movements. Ignoring these dynamic elements results in models that perform well only under specific, historical conditions, failing to deliver consistent returns in the face of the market’s inherent unpredictability.

Walk-forward analysis demonstrates that the strategy outperforms buy-and-hold across a range of <span class="katex-eq" data-katex-display="false"> heta</span> values, even without transaction costs, as evidenced by the cumulative trade count. — Walk-forward analysis demonstrates that the strategy outperforms buy-and-hold across a range of $heta$ values, even without transaction costs, as evidenced by the cumulative trade count.

Constructing Signals Rooted in Causal Reality

The Causal Signal Framework addresses the critical issue of look-ahead bias in algorithmic trading by constructing signals exclusively from past and present data. Look-ahead bias occurs when a trading signal inadvertently incorporates information not available at the time of decision-making, leading to unrealistically optimistic backtest results and poor performance in live trading. This framework ensures signals are generated using only data observable at the current time or in the historical record, preventing the use of future information that would invalidate real-time applicability. By strictly adhering to this principle, the Causal Signal Framework enables the creation of robust and reliable trading strategies suitable for automated, real-time implementation where instantaneous decision-making is paramount.

Causal Observables are foundational to constructing trading signals that avoid look-ahead bias. These variables are defined as those which can be fully determined using data available at the time of calculation, strictly prohibiting the inclusion of any future information. This is achieved by relying solely on past and present market data – such as open, high, low, close, and volume – and avoiding the use of indicators that require subsequent periods for their calculation. The construction of Causal Observables is critical for ensuring that trading strategies can be reliably implemented in real-time and accurately evaluated through backtesting, as they accurately reflect the information available to traders at any given moment.

The Causal Signal Framework utilizes three primary components for signal generation. Robust Normalization standardizes indicator values, mitigating the impact of differing scales and ensuring consistent weighting. Causal Smoothing techniques, such as time-weighted average price (TWAP) or exponential moving averages calculated solely on past data, are employed to reduce high-frequency noise and improve signal clarity. Finally, Derivative-Based Enhancement incorporates forward-looking information by calculating rates of change or differences of causal observables, providing insights into the momentum and potential direction of market movements without introducing future data dependencies.

Combining a signal with its scaled, causal derivative introduces a systematic phase advance, particularly around zero-crossings, and this effect can be adaptively controlled by modulating a gating parameter <span class="katex-eq" data-katex-display="false">\lambda_{2}</span>, suppressing derivative contributions away from zero-crossings while maintaining the derivative-enhanced signal when <span class="katex-eq" data-katex-display="false">\lambda_{2} = 0</span>. — Combining a signal with its scaled, causal derivative introduces a systematic phase advance, particularly around zero-crossings, and this effect can be adaptively controlled by modulating a gating parameter $\lambda_{2}$ , suppressing derivative contributions away from zero-crossings while maintaining the derivative-enhanced signal when $\lambda_{2} = 0$ .

A Composite Approach to Decoding Market Dynamics

The Composite Observable is constructed by integrating four distinct technical indicators – the Relative Strength Index (RSI), Money Flow Index (MFI), Moving Average Convergence Divergence (MACD), and Bollinger Bands Percentage (BB%) – to provide a consolidated assessment of market conditions. RSI measures the magnitude of recent price changes to evaluate overbought or oversold conditions, while MFI incorporates volume into this analysis. MACD identifies trend changes by comparing moving averages, and BB% indicates price positioning relative to Bollinger Bands, signaling potential breakouts or reversals. Combining these indicators aims to reduce the impact of false signals generated by any single indicator and offer a more nuanced understanding of price momentum, trend strength, and potential price extremes.

The integration of multiple technical indicators into a composite observable enhances the capture of diverse market signals by mitigating the limitations inherent in any single indicator. Individual indicators often exhibit biases or perform suboptimally under specific market conditions; combining them provides a more holistic assessment. For example, a momentum indicator like the MACD can be validated or refuted by a volume-based indicator like MFI, reducing the likelihood of false signals. This diversification of signal sources directly contributes to improved strategy robustness by decreasing sensitivity to the idiosyncrasies of individual indicators and increasing the probability of consistent performance across varying market environments. Consequently, strategies utilizing composite observables demonstrate a greater capacity to adapt to changing conditions and maintain profitability over time.

The Causal Signal Framework addresses a critical issue in technical analysis: spurious correlations between indicators and market movements. By establishing causal validity, the framework ensures that observed signals from composite observables – such as those combining RSI, MFI, MACD, and BB% – are genuinely predictive of future price action and not merely coincidental. This is achieved through rigorous backtesting methodologies and statistical analysis designed to identify and eliminate indicators that lack demonstrable causal links to market outcomes. The framework prioritizes signals originating from fundamental drivers, minimizing the risk of acting on false positives generated by random noise or unrelated market events, and ultimately improving the reliability of trading strategies.

The advanced composite indicator <span class="katex-eq" data-katex-display="false">F_0(t+6)</span> (blue) closely tracks the horizon-based trend indicator <span class="katex-eq" data-katex-display="false">\Delta^{(10)}(t)</span> (red), especially at zero-crossings that signify shifts in the underlying regime, as visualized by the corresponding regime markings in the lower panels. — The advanced composite indicator $F_0(t+6)$ (blue) closely tracks the horizon-based trend indicator $\Delta^{(10)}(t)$ (red), especially at zero-crossings that signify shifts in the underlying regime, as visualized by the corresponding regime markings in the lower panels.

Translating Signals into Disciplined Action

Decision functionals represent a critical evolution in automated trading, effectively converting raw market data into concrete investment strategies. These aren’t simply algorithms reacting to price changes; they are sophisticated systems designed to interpret a multitude of signals – volume, volatility, order book dynamics – and synthesize them into executable trade orders. The core function is translation: transforming abstract indicators of market sentiment into precise instructions for buying or selling assets. This process moves beyond basic technical analysis by allowing for the incorporation of complex, multi-faceted criteria, and ultimately enables a machine to mimic, and potentially surpass, the nuanced decision-making of a human trader. By streamlining the path from observation to execution, decision functionals minimize latency and maximize the potential for profit in fast-moving markets.

The implementation of a hysteresis-based decision rule represents a crucial mechanism for tempering reactivity in automated trading systems. This rule introduces a threshold that must be crossed before a trading signal triggers an action, effectively filtering out noise and preventing the system from overreacting to short-term market fluctuations. By demanding a sustained signal – a move beyond the initial threshold and remaining there – the rule minimizes whipsaws and the associated transaction costs. This lag between signal and action doesn’t diminish responsiveness to genuine trends; instead, it fosters stability and allows the system to discern meaningful price movements from random volatility, ultimately leading to more consistent and reliable performance. The effect is akin to a physical system with inertia, resisting abrupt changes and promoting smoother operation even amidst chaotic market conditions.

Decision functionals demonstrate considerable adaptability through the integration of horizon-based trend objectives, allowing for precise calibration to diverse investment strategies. Rather than reacting uniformly to all market shifts, these functionals can be programmed to prioritize trends aligning with a specific timeframe – be it short-term day trading, swing trading over weeks, or long-term investing spanning years. This nuanced approach involves weighting recent price movements more heavily for shorter horizons, and giving greater significance to long-term averages for extended investment goals. Consequently, the functional effectively filters noise and focuses on signals most relevant to the designated timeframe, thereby optimizing trade execution and enhancing portfolio performance aligned with the investor’s particular objectives. The ability to dynamically adjust responsiveness based on the intended holding period represents a significant advancement in automated trading systems, moving beyond generic responses toward genuinely tailored investment strategies.

Under a walk-forward protocol, hyperparameter selection frequencies reveal distinct preferences for EURUSDT and BTCUSDT, indicating differing optimal configurations for each trading pair.

Rigorous Validation and Deployment for Enduring Performance

Walk-forward selection offers a stringent evaluation of trading strategies by replicating the experience of live trading, but on historical data. Instead of simply backtesting a strategy across an entire dataset, this technique sequentially trains and tests the strategy on rolling windows of data, mimicking how a trader would adapt to new market conditions in real-time. The process begins by training the strategy on an initial period, then testing its performance on the subsequent out-of-sample period; this training/testing window then ‘walks forward’ through time, continuously retraining and re-evaluating. This approach significantly reduces the risk of overfitting – where a strategy performs well on historical data but fails in live trading – by assessing its ability to generalize to unseen data, providing a more realistic and reliable measure of its potential profitability and robustness.

A key challenge in developing algorithmic trading strategies lies in avoiding overfitting – a scenario where a strategy performs exceptionally well on historical data but fails to deliver results in live markets. Walk-forward selection directly addresses this by rigorously testing the strategy’s resilience. Instead of optimizing parameters on the entire dataset at once, the technique iteratively trains the strategy on a portion of the historical data, then tests its performance on unseen data. This process simulates real-world trading conditions more accurately, providing a realistic assessment of how the strategy will generalize to future, previously unobserved market dynamics. Consequently, successful performance during walk-forward analysis instills greater confidence in the strategy’s robustness and its potential for sustained profitability, minimizing the risk of deploying a system that is merely memorizing past patterns rather than capitalizing on genuine market signals.

The research details a causal signal engineering framework designed to actively manage portfolio risk, resulting in demonstrably smoother equity curves and tighter drawdown control without sacrificing overall performance. This approach goes beyond simple prediction; it reshapes the inherent risk profile of trading signals through a structured causal methodology. Analysis of both EURUSDT and BTCUSDT reveals a median position holding time of approximately 22-23 minutes, suggesting the strategy adeptly navigates a market environment characterized by both rapid, short-term fluctuations and the influence of more sustained, longer-term market regimes. This balance indicates the framework isn’t solely reliant on high-frequency scalping or exclusively focused on extended trend-following, but instead dynamically adapts to prevailing market conditions to optimize risk-adjusted returns.

Walk-forward optimization reveals that both EURUSDT and BTCUSDT exhibit right-skewed position durations, characterized by frequent, short-lived positions alongside a significant proportion of sustained regimes.

The pursuit of robust trading strategies, as detailed in the study, inherently confronts the challenge of non-stationary systems. The paper’s focus on causal inference and walk-forward optimization attempts to distill signal from noise, a process mirroring a fundamental existential search for meaning. As Søren Kierkegaard stated, “Life can only be understood backwards; but it must be lived forwards.” This aligns with the study’s methodology – learning from past market behavior (backward understanding) to adaptively construct strategies for future, uncertain conditions (forward action). The reduction of complexity through causal observables, a core tenet of the paper, echoes the sentiment that true understanding isn’t found in accumulating data, but in discerning essential relationships.

Beyond the Horizon

The presented work addresses, with a degree of pragmatic success, the persistent challenge of non-stationarity in financial systems. However, the very act of defining ‘causal observables’ introduces a structural dependency that merits further scrutiny. The selection of heterogeneous indicators, while empirically motivated, remains susceptible to the observer’s bias-a ghost in the machine of any predictive model. The illusion of control is comforting, yet the market, as a complex adaptive system, will inevitably reveal the limitations of any fixed causal structure.

Future iterations should not focus on expanding the indicator set, but on minimizing its influence. A truly robust system would ideally operate with a minimal, theoretically grounded set of variables-perhaps even a single, carefully constructed observable. The current reliance on walk-forward optimization, while effective for adaptation, is fundamentally reactive. The pursuit of genuinely anticipatory mechanisms-those that can discern structural shifts before they manifest in observable data-remains the central, and likely intractable, problem.

Emotion is a side effect of structure. The desire for ‘smooth equity dynamics’ and ‘drawdown control’ are merely expressions of a deeper aversion to uncertainty. Clarity is compassion for cognition. The ultimate test of this, and all similar, work will not be its profitability, but its parsimony-the degree to which it can explain complex behavior with the fewest possible assumptions.

Original article: https://arxiv.org/pdf/2603.13638.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/