The Vault

The End of Prediction: Proof from the Data

“When the illusion of prediction breaks, the process begins.”

The Premise

For decades, finance has been built on the comforting belief that markets are predictable if we can only find the right model. Economists call this faith in stability stationarity, the idea that market relationships are fixed, correlations are dependable, and volatility is a mild inconvenience around an otherwise stable mean. Reality tells a different story. Markets are non-stationary, reflexive, and adaptive. Their participants learn, act, and respond to one another’s actions. Every attempt at prediction alters the system being predicted. The outcome is not true randomness but something more subtle and far more complex: structured unpredictability.

The following analysis examines seven empirical exhibits drawn from the E-mini S&P 500 futures market (ticker: ES), representing the electronically traded S&P 500 index futures contract listed on the Chicago Mercantile Exchange (CME). This continuous Panama-adjusted series serves as a robust proxy for U.S. equity behaviour. The ES contract was chosen for its depth, liquidity, and diversity of participants, making it an ideal setting for observing adaptation and feedback in action.

While this analysis focuses on ES, the findings are replicable across any sufficiently liquid futures market. The principles of non-stationarity, correlation decay, and adaptive feedback are not unique to equities; they are structural characteristics of all markets shaped by human and algorithmic interaction. Wherever liquidity concentrates and participants adapt to one another, prediction decays and process prevails.

Each exhibit visualises a different form of instability: correlation decay, parameter drift, volatility clustering, or feedback-driven reflexivity. Together, they tell a single story:

Prediction dies where adaptation begins. Process, not forecast, is the only sustainable response.

Why Prediction Fails in Adaptive Markets

Before diving into the data, it helps to understand why prediction repeatedly disappoints. Economists often model markets as if they were closed systems governed by static laws. But financial markets are complex adaptive systems, and they differ in four key ways.

Non-Stationarity

The statistical properties of markets change over time. Mean returns, variances, and correlations are not constants. They evolve as regimes shift, liquidity flows, and participants adapt. Models calibrated on one period decay in the next.

Correlation Instability

Relationships between assets, or even between lagged versions of the same asset, are transient. When traders act on a correlation, they compress or reverse it. This self-negating property ensures that predictive edges have a short shelf life.

Volatility Clustering

Markets breathe. Periods of quiet are followed by storms, and storms give way to calm. The very presence of volatility clustering shows that variance is path-dependent, a product of memory and feedback rather than a random scatter.

Fat Tails

Extreme events occur far more often than Gaussian theory predicts. These fat tails arise from feedback loops, leverage, and herding, all natural consequences of reflexivity in an interconnected system. Each of these features undermines the assumption of equilibrium. The market is not a tidy probabilistic machine. It is a living, learning ecosystem that reacts to the very models designed to describe it.

The Empirical Proof

Using 40 years of Panama Adjusted real daily ES data aggregated to month-end closes, we translated these theoretical claims into measurable evidence. The data series runs across multiple decades, capturing both calm periods and crises. The exhibits below demonstrate how the elegant assumptions of predictability break down in practice. Exhibit 1 – Rolling 60-Month Coefficient of Determination: Predictability Decay: Rolling 60-month Coefficient of Determination fluctuates around zero, with short-lived peaks in 1996 and 2009 that coincide with phases of collective alignment. Each spike fades as heterogeneity and adaptation return. What we measured We regressed each month’s return on the previous month’s return using rolling 60-month windows and recorded the coefficient of determination, the percentage of variation in the current month’s return that could be statistically explained by the prior month. This shows how much of next month’s behaviour could, at any point in time, be “predicted” by what happened last month. What we observed The line stays close to zero through most of the record, reflecting a persistent absence of linear predictability. Yet there are two noticeable spikes that briefly rise above the noise: one around 1996 and another in 2009. In 1996, the coefficient climbed toward twelve percent during the sustained expansion that preceded the late-1990s technology boom. Market enthusiasm, easy monetary policy, and the early spread of online trading encouraged one-way positioning and momentum reinforcement. The 2009 spike, which approached fifteen percent, coincided with the reflexive rebound that followed the Global Financial Crisis. Extraordinary policy interventions, quantitative easing, and broad “risk-on” behaviour drove collective movement in a single direction. These phases mark short periods when diversity of behaviour compressed and participants moved in phase. As liquidity stabilised and market heterogeneity returned, the coefficient fell back toward zero. The brief alignment dissolved as new entrants, changing expectations, and adaptive strategies restored noise to the flow of returns. Interpretation These spikes are the statistical fingerprints of synchronisation. They appear when market participants temporarily behave as a single organism, either through optimism that fuels a rally or through collective reaction to systemic stress. The 1996 episode captured the build-up of positive feedback that later culminated in the tech-bubble collapse. The 2009 episode reflected the forced coherence of post-crisis recovery, when policy and psychology aligned in rebuilding confidence. In each case, predictability appeared not because the market became orderly but because it briefly lost diversity. When those feedback loops broke and behaviour diversified again, predictability vanished. This pattern is evidence of an adaptive system that self-organises, synchronises, and then resets. Key insight: Predictability emerges only when variety disappears. Once adaptation resumes, the edge evaporates and the market returns to noise.

Exhibit 2 – Rolling 60-Month Slope Coefficient: Parameter Drift: The slope coefficient alternates between positive and negative regimes. Negative phases (1992–1996 and 2009–2020) coincide with consolidation, mean reversion, and volatility suppression. Positive phases (1996–2009) align with expansionary periods driven by liquidity, leverage, and crowd synchrony. What we measured Using the same regression framework as Exhibit 1, we tracked the slope coefficient of monthly returns regressed on the prior month’s return. This slope captures how the current month’s price behaviour relates to the previous month. Positive values indicate trend reinforcement (momentum), while negative values reflect corrective behaviour (mean reversion). Tracking these changes over time allows us to see how the underlying structure of market behaviour evolves. What we observed The slope coefficient alternates between positive and negative regimes that align closely with major market phases.

1992 to 1996: The slope remained negative, marking a period of consolidation and recovery following the 1987 crash and the early 1990s recession. Liquidity was still rebuilding, and institutional relative-value trading reinforced mean-reverting tendencies.
1996 to 2009: The slope turned positive for an extended period, capturing a long era of trend reinforcement. This phase included the technology boom, the credit expansion that followed, and the post-crisis rebound after 2008. Liquidity growth, the rise of passive flows, and increasing leverage created sustained directional persistence. Markets trended as participants acted in unison, reinforcing rather than fading momentum.
2009 to 2020: The slope again turned negative and stayed there for much of the following decade. This period reflected the dampening effects of central bank intervention and the growing influence of algorithmic trading. Volatility was suppressed, trends were shorter-lived, and crowd behaviour became reflexive and self-correcting.

Through these alternating regimes, the slope coefficient rarely held steady for long. It rose when behaviour aligned and fell when diversity returned. Interpretation The drifting slope illustrates the adaptive feedback at the heart of market dynamics. When collective models and beliefs align, crowd behaviour amplifies trends, creating positive feedback. When that synchrony overextends, contrarian forces, policy shifts, or structural breaks push the system into correction, reintroducing negative feedback. This continual back-and-forth between reinforcement and reversion is the market’s natural rhythm. It shows that price behaviour is not stationary but constantly shaped by the interaction of agents who adapt to each other’s actions. Each swing in the slope marks a reconfiguration of the system, from expansionary alignment to corrective diversity and back again. Key insight: Parameter drift is the visible trace of adaptation. Positive phases reflect crowd alignment and trend reinforcement. Negative phases reveal reflexivity, self-correction, and renewed diversity. The market never stays fixed; it continually evolves.

Exhibit 3 – Volatility Clustering & Regime Shifts: Monthly realised volatility, with 12-month and 36-month moving averages, shows pronounced clusters around 1998 (LTCM), 2000–2003 (dot-com collapse), 2008–2009 (GFC), and 2020 (pandemic shock). Calm phases between these events reflect temporary periods of crowd synchrony and suppressed diversity that later give way to renewed turbulence. What we measured Using daily ES data, we calculated monthly realised volatility as the standard deviation of daily returns within each month. We then smoothed this series with 12-month and 36-month moving averages to highlight periods where volatility persisted above or below its long-term mean. The goal was to visualise how volatility behaves through time and whether it is random or exhibits persistence and memory. What we observed Volatility does not distribute evenly through time. It clusters. Periods of quiet stability are followed by long sequences of turbulence, and those clusters often align with moments of structural transition in the market.

1998: A sharp rise in realised volatility marks the Long-Term Capital Management (LTCM) crisis, when leveraged arbitrage positions unwound under liquidity stress. This event exposed the fragility of tightly coupled strategies and produced a temporary surge in feedback that rippled across global markets.
2000 to 2003: Another extended rise aligns with the dot-com collapse, as technology valuations unwound and risk tolerance fell. Volatility remained elevated for several years, showing how shocks persist as behavioural aftereffects rather than isolated spikes.
2008 to 2009: The most pronounced cluster coincides with the Global Financial Crisis, when systemic deleveraging and policy intervention created a cascade of feedback loops that engulfed every asset class.
2020: The COVID-19 pandemic generated the sharpest short-term volatility spike on record, as global uncertainty compressed reaction times and forced simultaneous liquidation across markets.

Between these crises, volatility repeatedly settled into extended low-volatility phases such as 2004–2006 and 2012–2018. These quieter intervals reflected periods of crowd synchrony and liquidity expansion that later gave way to renewed turbulence. The alternating pattern between high and low regimes forms the essential rhythm of the market: stability creates the conditions for instability, and instability restores diversity. Interpretation Volatility clustering is not random noise. It is a sign of memory within the system, showing that market behaviour retains traces of its past. When volatility rises, risk controls tighten, liquidity providers withdraw, and trend followers amplify directional moves, all reinforcing the expansion. As participants adapt, their reactions generate secondary waves that prolong the high-volatility state. Conversely, long calm phases occur when participants collectively suppress volatility through convergent behaviour, narrow positioning, and confidence in equilibrium. These low-volatility states are themselves unstable because they reduce diversity and prepare the ground for the next expansion in volatility. This pattern of persistence and clustering is exactly what we expect from a complex adaptive system: feedback-driven, state-dependent, and path-sensitive. Volatility does not revert to a constant mean. It evolves over time, reflecting the ongoing interplay between amplification and suppression. Key insight: Volatility is not the absence of order. It is the memory of past order breaking down. Each cluster records the system’s effort to adapt and re-balance after synchronisation collapses.

Exhibit 4 –The Stability Trap: Out-of-Sample Collapse of Predictive Models: Comparison of predictive and reactive systems using two consecutive validation periods. The predictive model’s out-of-sample performance collapses toward zero after each recalibration, revealing the fragility of optimisation. The reactive trend process maintains positive, stable results across regimes, demonstrating robustness through adaptation rather than prediction. What we measured We compared a simple predictive model with a reactive breakout process using rolling 60-month training and testing windows. Each model was recalibrated every five years and then applied to the next five-year period. The chart below displays only the out-of-sample results for each of these validation periods. For each window, we calculated a Sharpe-style measure (mean monthly return divided by standard deviation). This statistic was used purely as a comparative gauge of structural stability, not as a measure of skill or performance quality. Why this distinction matters The Sharpe ratio assumes a stationary world with constant volatility and normally distributed returns. Financial markets are neither. When regime shifts and fat tails dominate, this ratio can be misleading. In this analysis it is used only to illustrate how predictive systems deteriorate when their environment changes, while reactive systems retain coherence. What we observed Both Forecast_First and Forecast_Second represent the predictive model’s out-of-sample performance following two separate calibration periods. In each case, the Sharpe-style reading is near zero, showing that any apparent in-sample precision vanished once new data arrived. The model’s relationships were not stable through time. The Breakout_First and Breakout_Second bars correspond to the reactive trend process over the same validation windows. These results remain positive and consistent, showing that the breakout method adapts across regimes. Its parameters were fixed, and its logic responded only to price movement rather than attempting to forecast it. Interpretation This chart visualises the stability trap of predictive design. Models that are finely tuned to past data appear precise within their calibration window but collapse when confronted with structural change. Optimisation creates fragility because it locks a model into a transient pattern. Once the market transitions, the previously “optimal” settings become irrelevant. Reactive systems avoid this fate. They do not rely on the persistence of relationships; they depend on observation and response. The breakout model in this analysis did not require recalibration or parameter adjustment. Its steady out-of-sample readings demonstrate that adaptability is achieved through simplicity, not prediction. Key insight: The chart shows only out-of-sample results. The predictive model’s near-zero performance illustrates how its in-sample precision evaporates under changing conditions. The breakout model’s stability confirms that reaction, not prediction, is the path to durability.

Exhibit 5 – Forecast vs Breakout: Cumulative Payoffs: Cumulative return comparison of forecast and breakout systems. The forecast model oscillates around zero, losing accumulated gains after each structural shift. The breakout model compounds through time with irregular but persistent growth, reflecting the power of reactivity over prediction in non-stationary markets. What we measured We compared two simplified systems applied to the same monthly ES data.

A forecast model, which uses historical correlations and mean reversion logic to predict the next month’s return direction.
A reactive breakout model, which ignores prediction and simply enters when price closes above or below a recent range.

Both models were normalised for volatility and trade frequency, so differences in outcome reflected only behavioural structure. Each model’s cumulative return was tracked over time. What we observed The results could not be more distinct. The forecast model performed well during stable or mean-reverting environments but consistently decayed after structural shifts. Gains accumulated in quiet phases, only to be erased during trend expansions or crisis transitions. Over the full sample, the forecast curve oscillated around zero with no enduring edge. The breakout model, by contrast, showed irregular but sustained growth. It endured long flat periods when trends were scarce, yet captured the large moves that followed regime transitions. Its cumulative curve rose unevenly but persistently, showing that payoffs arrived in bursts rather than in a smooth or predictable stream. This divergence demonstrates a fundamental principle: prediction depends on continuity of structure, while reaction depends only on change. The forecast approach requires the future to behave like the past; the breakout approach profits precisely when that assumption fails. Interpretation These results reveal the asymmetry between predictive fragility and reactive resilience. Predictive strategies thrive on short-term stability but collapse when the underlying process shifts. Reactive trend systems underperform during stability but excel when structure breaks. Over time, the reactive approach compounds gains from rare but powerful transitions, while the predictive model oscillates within the confines of noise. This is not a story of efficiency or skill; it is a reflection of how markets organise themselves. Systems anchored in prediction chase a vanishing equilibrium. Systems grounded in reaction align with the market’s adaptive flow. Key insight: Forecasting extracts comfort from continuity. Breakouts extract profit from change. The former depends on knowing, the latter on surviving.

Exhibit 6 – Reflexive Crowding Proxy: Scatterplot of model alignment versus next-month portfolio returns. The regression line is nearly flat, confirming that no stable relationship exists between collective alignment and subsequent payoff. This reflects the self-neutralising nature of reflexive markets. What we measured We examined how collective positioning affects future payoffs by creating an alignment metric that captures how often the ensemble of systems pointed in the same direction. We then compared this alignment with next-month portfolio returns using a simple least-squares regression and confirmed the findings with a rank correlation test. What we observed The scatter of points forms a broad, featureless cloud with a regression line that is nearly flat. In practical terms, there is no statistically meaningful relationship between alignment and subsequent return. Periods of extreme crowding occasionally precede weaker outcomes, but the overall tendency is neutral. This lack of slope is itself informative: the system adapts quickly enough that any transient link between consensus and payoff is erased. Interpretation The near-zero slope illustrates how feedback and adaptation extinguish stable relationships. When models converge, their collective behaviour reshapes market structure, cancelling the very edges they sought to exploit. Reflexivity ensures that no simple rule linking alignment to payoff can persist for long. This is not a failure of modelling but an expression of the market’s self-regulating nature. Once participants act on a perceived advantage, the system evolves to absorb it, leaving behind statistical flatness, the hallmark of adaptation. Key insight: The absence of correlation is not noise. It is the signature of a self-correcting system that erases its own footprints.

Exhibit 7 – Edge Half-Life: Correlation Decay: Correlation between model signals and realised returns at increasing lags. Correlation decays rapidly, stabilising near zero within a year. The short half-life of informational edges illustrates market adaptation. The enduring edge of trend following arises not from prediction, but from participation in the universal dynamics that create Outliers. What we measured To explore how information and edge decay through time, we calculated the correlation between model signals and realised returns at increasing time lags. Each lag represents how long a signal remains relevant before the relationship weakens. The point where correlation falls below half of its initial value defines the edge half-life. We repeated this process across multiple systems and parameter configurations to observe the consistency of decay. What we observed The correlations between model signals and realised returns decline rapidly. Within a few months, most relationships lose more than half of their initial strength. By twelve months, correlations fluctuate around zero, indicating that no lasting relationship persists between past and future. Occasionally, a system will show short-lived persistence, but these instances are transient and inconsistent across regimes. The overall pattern reveals steep decay followed by random oscillation. This finding reinforces the earlier exhibits: structure appears, interacts, and then dissolves as adaptation absorbs it. Interpretation The short half-life of correlation demonstrates how quickly informational edges are consumed by the market. Once a signal becomes widely recognised, it is arbitraged away. Participants act on it, their actions change the environment, and the edge vanishes. This is not random behaviour but the defining feature of an adaptive system. Every participant alters the structure they depend on, ensuring that most discovered relationships self-destruct through reflexive feedback. Many readers might ask, “If all edges decay, why does trend following continue to work?” The answer lies in the source of the edge. Predictive edges rely on stable correlations or repeatable signals. They depend on the world behaving tomorrow as it did yesterday. Trend following does not. Its edge is not derived from prediction but from process. It captures the emergent behaviour that arises when a complex adaptive system departs from equilibrium and generates directional movement. These moves are not forecastable, but they are inevitable outcomes of feedback and flow. Trend following’s persistence is therefore not a statistical anomaly but a structural necessity. Markets are fractal systems that expand and contract through feedback. They create trends of varying scale as they transition between order and disorder. These transitions produce asymmetric opportunities, or Outliers, that are universal in adaptive systems. Predictive edges vanish because they chase constancy. The trend following edge survives because it aligns with change. It does not assume stability; it thrives on instability. It does not seek to know; it seeks to respond. Its robustness comes not from precision, but from its acceptance of uncertainty and its capacity to adapt to it. Key insight: Predictive edges decay because they rely on permanence in a system that evolves. Trend following endures because it aligns with the fractal rhythm of change. Its edge is not a signal to discover but a principle to obey.

Synthesis: The End of Prediction

Across these seven exhibits, the data tell a clear and consistent story. Every measurable form of predictability: correlations, parameters, regressions, and informational edges, decays through time. Relationships that appear stable in one regime collapse in the next. Patterns that once offered advantage are absorbed and neutralised by adaptation. This is not market inefficiency; it is the market’s defining feature. The financial system behaves as a living organism, constantly reshaping itself in response to the behaviour of its participants. Each attempt to model it alters its structure, and every successful discovery accelerates its own decay. The result is a market that appears noisy, but whose apparent randomness is simply the visible expression of continual adaptation. The evidence is unambiguous:

Predictability decays. Relationships such as return autocorrelation rise briefly and then vanish.
Parameters drift. The slope of reinforcement alternates between positive and negative as crowd behaviour flips between expansion and correction.
Volatility clusters. Stability and instability follow each other in rhythmic succession as feedback loops compress and release.
Predictive models fail out-of-sample. What fits yesterday’s structure cannot survive tomorrow’s transition.
Crowding neutralises advantage. As consensus grows, opportunity shrinks.
Informational edges die fast. Correlations collapse within months as the system rebalances itself.

If all edges decay, what remains? The answer is the edge that does not depend on prediction. Trend following does not rely on repeatable signals or stable parameters. Its edge arises from the way complex adaptive systems move through positive feedback, negative feedback, and noise. These forces are universal. They are found in weather systems, ecosystems, economies, and markets alike. Trends are not anomalies. They are the signatures of an adaptive world reorganising itself. They emerge when balance breaks, when energy or liquidity shifts from one state to another. These transitions cannot be forecast, yet they are inevitable. The role of the trend follower is not to predict them but to recognise and participate in them when they occur. The persistence of trend following’s edge is therefore not a statistical accident but a structural consequence of how the universe evolves. Fractals expand through feedback, and Outliers arise naturally within them. To align with that process is not to exploit a temporary inefficiency but to act in accordance with a universal law of motion: that change, not stability, drives all progress. Final insight: Prediction seeks comfort in repetition. Trend following seeks truth in transformation. What decays for the forecaster endures for the observer who accepts uncertainty and rides the wave of adaptation.

The Fragility of Regime Filters

Some may argue that our analysis does not invalidate prediction entirely, but instead reveals that predictive systems work during certain favourable regimes. With enough sophistication, regime filters, volatility state identifiers, or machine learning methods such as Hidden Markov Models, it might seem possible to capture these temporary windows of predictability and exploit them for profit. There is truth in this observation. Markets do cycle through phases that appear more predictable than others. However, the advantage gained from identifying these regimes is always conditional and transient. The deeper issue lies not in exploiting a temporary edge, but in what happens when that regime ends. Predictive frameworks that rely on regime classification must commit to structural assumptions about what defines a favourable environment. These assumptions create fragility. When the market transitions, the same filters that once protected the model now trap it within a decaying structure. The very act of optimisation toward a specific regime embeds vulnerability to its failure. This is the hidden cost of precision. Models designed to excel in particular conditions often do so by overfitting to those very conditions. When the environment shifts, as it inevitably does, they fail asymmetrically. Gains earned slowly in stable periods are erased violently when feedback reverses. The result is the familiar pattern of negative skew: long stretches of apparent success punctuated by catastrophic breakdowns. The problem, therefore, is not in recognising that predictive tendencies can exist, but in believing they can be captured safely or sustainably. To exploit a fleeting inefficiency requires constraining flexibility, and in doing so, the model becomes brittle. Its survival depends on the persistence of a regime that cannot be forecast. Reactive systems avoid this trap by refusing to define the environment in advance. They accept that structure will change and that no model can reliably anticipate when or how. Their robustness comes not from predicting the next regime but from adapting to it when it arrives. Key insight: Regime filters create comfort through structure but fragility through assumption. In adaptive markets, the safer path is not to predict the next phase, but to remain free to respond to any.

Conclusion: Process Over Prediction

Markets are not random, yet they are not predictable either. They are deterministic systems whose behaviour unfolds through feedback and adaptation. Their movements arise from countless interactions among agents, each following rules that collectively exceed the reach of any model. Predictive systems assume continuity, but real markets deliver discontinuity. Forecasting assumes stationarity, but real markets deliver evolution. Edges based on correlation, optimisation, or fitted rules all decay because they rely on stability in a system defined by change. Across every exhibit, the evidence was clear. Relationships break down, parameters drift, volatility clusters, and informational edges erode. The more a model learns from the past, the less relevant that learning becomes. This is not noise or randomness; it is adaptation. The market constantly rewrites its own code in response to its participants. To thrive in such an environment requires abandoning prediction as the goal and embracing reaction as the process. Trend following, breakout systems, and other divergent frameworks succeed because they are aligned with the geometry of uncertainty. They do not depend on stable relationships or fixed assumptions. They respond to feedback, adapt to flow, and compound through the asymmetry of Outliers that arise when equilibrium breaks. Trend following’s edge persists because it is not a predictive edge. It is structural. It emerges from the same fractal rhythm that governs all adaptive systems, the universal alternation between stability and instability, feedback and reset. Its foundation is not in the repetition of patterns but in the inevitability of change. Final Thought Prediction is fragile because it requires the world to hold still. Process is robust because it accepts that it never will.

Epilogue: The End of Prediction is the Beginning of Understanding

The desire to predict is a natural impulse. It reflects our need for certainty in a world that refuses to stand still. Yet prediction belongs to a closed system, one where cause and effect can be contained within fixed boundaries. Markets are not such systems. They are open, adaptive, and alive. The more we observe, the more we influence. Every forecast becomes an input to the system it seeks to describe. Every discovery alters the path of those who act upon it. In this way, prediction and adaptation are inseparable, each reshaping the other in an endless loop of learning and forgetting. To understand markets, we must release the idea of control and embrace the process itself. The goal is not to foresee outcomes but to participate with awareness. The trader who reacts instead of predicts is not surrendering to chaos but aligning with the rhythm of the world as it truly is. Trend following represents this philosophy in practice. It does not claim to know what comes next. It listens, waits, and responds to the information contained in movement itself. In doing so, it turns uncertainty into structure and adapts alongside the market it observes. When we stop trying to predict, we begin to see. We discover that the market is not a puzzle to be solved but a living process to be understood: a mirror of adaptation, complexity, and the timeless pattern of change. Final Reflection: Prediction seeks to define the world. Understanding learns to move with it.

Appendix: Notes on Methods and Literature

Methodology Summary All analyses were performed on monthly data derived from continuous, Panama-style back-adjusted futures series to avoid artificial discontinuities at contract rolls. Each test used standard least-squares regression or correlation methods applied through rolling windows to capture the dynamic evolution of relationships through time.

Rolling coefficient of determination: Calculated by regressing each month’s return on the previous month’s return using a 60-month moving window. This measured the proportion of variation in next-month returns that could be explained by the prior month.
Rolling slope analysis: The same regression windows were used to examine the direction and magnitude of parameter drift across time, highlighting when crowd behaviour flipped sign.
Volatility clustering: Estimated through the rolling standard deviation of returns and the serial correlation of squared returns.
Reflexive crowding proxy: Constructed from the alignment frequency of ensemble model signals against realised portfolio returns.
Edge half-life: Computed by correlating model signals with realised returns at increasing lags; the half-life was the point where correlation fell below 50 percent of its initial value.

These measures are straightforward and reproducible in any statistical or spreadsheet environment. The goal was not to design a predictive model but to visualise the transience of relationships that predictive systems depend upon.

Supporting Literature

The findings of rapid correlation decay, parameter instability, and non-stationarity are well-established in financial research. Similar conclusions have been documented for decades:

Mandelbrot (1963, 1997): Demonstrated that price data are fractal and non-Gaussian, exhibiting self-similarity and volatility clustering.
Lo (1991): Found weak and unstable autocorrelation in returns, consistent with adaptive-market behaviour.
Farmer and Joshi (2002): Modelled markets as evolving ecosystems where arbitrage opportunities are continually created and destroyed by adaptive agents.
Peters (1994): Introduced fractal market theory, showing that multi-scale feedback processes explain both persistence and sudden breakdowns in structure.
Timmermann (2006): Reviewed empirical evidence for parameter instability and model breakdowns in forecasting, concluding that time-varying dynamics dominate financial series.

The tests presented here therefore confirm, rather than replace, a long body of academic evidence: predictive relationships in financial markets are fleeting because the system itself learns, adapts, and reorganises. What changes over time is not the presence of feedback, but its form.

Closing Reflection

These findings complete the circle of observation. The data confirm what experience already teaches: that the search for prediction in complex systems leads not to mastery but to movement. Every attempt to fix the future reshapes the present, and every discovery of order carries within it the seed of its own decay. To understand markets is not to forecast them, but to witness their perpetual act of becoming. In that awareness, process replaces prediction, and participation replaces control. The trader who accepts this truth does not surrender to uncertainty, they learn to move with it.