Randomness in Sequential Price DirectionNVIDIA CorporationBATS:NVDATrendAdvantageFigure 1. Example of Runs Test Z-score classifications. Grey regions indicate statistically Random-Like directional sequences, blue regions indicate Persistent directional behavior, and orange regions indicate Oscillatory (alternating) directional behavior. ## Objective An investigation was conducted to evaluate whether sequential **price direction** is governed by a random process. Specifically, directional price sequences were tested for randomness across **Daily, 4-hour, 1-hour, and 30-minute** timeframes. In addition, a robustness study was conducted to determine whether the results remained stable across Runs Test lookback windows of **30, 60, and 100 bars**. Randomness was evaluated using the **Wald–Wolfowitz Runs Test Z-Score** applied to rolling windows of close-to-close directional price changes. The objective of this study was **not** to predict future returns, but rather to quantify the statistical structure of directional price sequences. --- # Data The study consisted primarily of large-cap U.S. equities drawn from the S&P 100 and spanning multiple industry sectors. Two complementary datasets were analyzed. ### Timeframe Study * 107 U.S. equities * Daily, 4-hour, 1-hour, and 30-minute charts * Analysis period: **January 2020 – March 2026** **107 stocks × 4 timeframes = 428 independent price series** ### Lookback Robustness Study * 50 U.S. equities * Daily, 4-hour, 1-hour, and 30-minute charts * Runs Test lookback windows: * 30 bars * 60 bars * 100 bars **50 stocks × 4 timeframes × 3 lookbacks = 600 independent configurations** --- # Runs Test Z-Score Classification For every rolling window, the Runs Test Z-Score was calculated using close-to-close directional price changes and classified as follows. | State | Condition | | :---------- | :--------------- | | Persistent | Z < -1.96 | | Random-Like | -1.96 ≤ Z ≤ 1.96 | | Oscillatory | Z > +1.96 | For every stock, timeframe, and lookback window, the percentage of rolling windows falling into each classification was calculated. --- # Statistical Framework The primary endpoint was the **Random-Like percentage**. The following statistical procedures were applied: 1. Friedman Test (repeated-measures comparison across timeframes and lookback periods) 2. Kendall's W (effect size) 3. Wilcoxon Signed-Rank Tests with Holm correction (post-hoc comparisons) 4. Page Trend Test for the ordered hypotheses: ``` Daily < 4-hour < 1-hour < 30-minute 30-bar > 60-bar > 100-bar ``` --- # Primary Findings **Across all securities, timeframes, and lookback windows, approximately 95% of rolling windows were classified as Random-Like by the Wald–Wolfowitz Runs Test.** This remarkable level of stability persisted across every configuration examined. --- ## Timeframe Study (107 Stocks) Median Random-Like percentages were: | Timeframe | Random-Like (%) | | :-------- | --------------: | | Daily | 95.21 | | 4-hour | 95.10 | | 1-hour | 95.14 | | 30-minute | 94.99 | **Range across all four timeframes:** **0.22 percentage points** Despite analyzing more than four hundred independent price series, the measured level of directional randomness remained essentially unchanged across observation frequencies. --- ## Lookback Robustness Study (50 Stocks) Changing the Runs Test lookback window from 30 to 60 and 100 bars produced similarly small changes. | Lookback | Mean Random-Like (%) | | :------- | -------------------: | | 30 | 94.77 | | 60 | 94.04 | | 100 | 93.34 | **Total change:** **1.43 percentage points** Even the largest observed difference represented less than a **2% absolute change** in Random-Like classification. Across all tested configurations, Random-Like classifications consistently remained close to 95%. --- # Statistical Confirmation Repeated-measures statistical testing confirmed that these small numerical differences rarely translated into meaningful practical effects. ## Timeframe Study **Zero-Split Partition** * Friedman χ² = 3.056 * p = 0.383 * Kendall's W = 0.0095 Kendall's W indicates a **negligible effect size**, demonstrating that differences between timeframes were statistically trivial. The Page Trend Test found **no evidence** that directional randomness increases monotonically as timeframe decreases. --- ## Lookback Robustness Across all four timeframes, the largest observed effect size was: **Kendall's W = 0.281** Most comparisons produced considerably smaller effect sizes. These results indicate that increasing the Runs Test lookback window from 30 to 100 bars produces only minor changes in classification percentages. --- ## Summary Results ### Daily | Lookback | Persistent % | Random-Like % | Oscillatory % | |---|---:|---:|---:| | 30 | 2.36 | **94.77** | 2.87 | | 60 | 2.54 | **94.04** | 3.42 | | 100 | 2.33 | **93.34** | 4.33 | ### 4-Hour | Lookback | Persistent % | Random-Like % | Oscillatory % | |---|---:|---:|---:| | 30 | 1.96 | **95.31** | 2.73 | | 60 | 1.78 | **95.27** | 2.95 | | 100 | 1.61 | **94.79** | 3.60 | ### 1-Hour | Lookback | Persistent % | Random-Like % | Oscillatory % | |---|---:|---:|---:| | 30 | 2.47 | **95.17** | 2.36 | | 60 | 2.42 | **95.11** | 2.46 | | 100 | 2.43 | **95.15** | 2.42 | ### 30-Minute | Lookback | Persistent % | Random-Like % | Oscillatory % | |---|---:|---:|---:| | 30 | 2.53 | **95.04** | 2.43 | | 60 | 2.38 | **94.90** | 2.71 | | 100 | 2.38 | **94.75** | 2.87 | **P = Persistent R = Random-Like O = Oscillatory** The 100-bar Daily configuration exhibited a modestly lower Random-Like percentage (93.3%) and a correspondingly higher Oscillatory classification (4.3%). Although this suggests slightly more detectable directional structure at longer observation windows, the dominant classification remained overwhelmingly Random-Like. --- # Conclusions The combined studies produced one clear empirical result. **Across all datasets, approximately 95% of rolling windows were classified as Random-Like by the Wald–Wolfowitz Runs Test, regardless of whether prices were sampled on Daily, 4-hour, 1-hour, or 30-minute charts.** Furthermore, this conclusion remained remarkably stable across Runs Test lookback windows of **30, 60, and 100 bars**. --- # Key Takeaways ## 1. Directional price sequences are overwhelmingly Random-Like Approximately **95%** of all rolling windows were classified as Random-Like regardless of timeframe or lookback window. **Implication** Trading methodologies that rely primarily on directional price sequences are operating in a domain that was statistically indistinguishable from randomness for approximately 95% of the rolling windows examined. Visual chart patterns such as flags, engulfing candles, and morning stars therefore require rigorous out-of-sample validation before being considered evidence of a genuine trading edge. --- ## 2. Timeframe has little influence on directional randomness The hypothesized Adaptive Market Hypothesis ordering—that lower timeframes would exhibit progressively greater randomness—was not supported. Random-Like classifications remained remarkably uniform across all timeframes. **Implication** Changing chart timeframe is unlikely to reduce directional randomness. A trader moving from Daily to 30-minute charts should not expect materially different directional behavior. --- ## 3. Most of the informative structure resides in the minority tail states Persistent and Oscillatory classifications together accounted for only about **5%** of all rolling windows. These are the only regions where the Runs Test rejects the null hypothesis of randomness, indicating statistically detectable departures from random directional sequencing. **Implication** For quantitative traders using regime filters, the Random-Like state may be viewed as a reduced-confidence environment, while Persistent and Oscillatory classifications deserve greater analytical attention. --- ## 4. Longer lookback windows show a modest increase in Oscillatory classifications At the 100-bar lookback, Oscillatory classifications increased modestly, particularly on Daily and 4-hour charts. **Implication** Whether this behavior reflects exploitable mean-reverting dynamics requires separate predictive validation. The Runs Test alone cannot establish predictive trading value. --- ## 5. Candlestick and price-action methodologies warrant independent statistical validation Many candlestick and price-action methodologies derive their signals from directional price sequences—the same domain found to be Random-Like for approximately **95%** of the rolling windows analyzed. **Implication** Pattern recognition alone should not be regarded as sufficient evidence of predictive power. Any claimed trading edge should be demonstrated through independent statistical testing against an appropriate randomness baseline. --- # Conclusion This study suggests that **directional sign information alone contains limited statistical structure across standard trading timeframes.** Future research may therefore benefit from incorporating additional information beyond directional sign, including return magnitude, volatility, trading volume, or cross-asset relationships, which may contain structural information not captured by directional sequencing alone. ## What This Study Does Not Say To prevent a common misinterpretation, a 95% Random-Like classification does **not** mean that profitable trading is impossible, nor does it imply that markets are informationally efficient. It means only that the raw sequence of positive and negative price changes—the ordering of green and red bars—contains **limited statistical information when considered in isolation**. A directional sequence that appears random under the Runs Test can still coexist with genuine market inefficiencies. This study does not rule out profitability; rather, it suggests that **directional sign alone is unlikely to be a sufficient source of trading edge**. Nor does it dismiss the importance of return magnitude. A price series can exhibit near-random directional sequencing while remaining highly structured in the size of its price movements. Trend-following strategies that capture relatively rare but outsized moves, and mean-reversion strategies that exploit predictable volatility behavior, may retain a durable edge even when the sequence of winning and losing trades resembles a coin flip. Likewise, this study does not dismiss non-price information. Trading volume, order flow, liquidity conditions, macroeconomic influences, and cross-asset relationships may all contain structural information that binary directional sign analysis intentionally ignores. Finally, the approximately 5% of windows classified as Persistent or Oscillatory should not be viewed as trivial. In quantitative finance, relatively small statistical edges can have meaningful practical value when exploited consistently. Casinos operate on a house edge of 1–5% depending on the game, and generate consistent, predictable, compounding revenue.**If an algorithm can reliably identify these departures from randomness in real time, they may represent informative market regimes worthy of further investigation.**