Statistical Astrology: The Use of Irregular Seasonality in Quantitative Time Series Analysis and Forecasting
The research itself has taken the form of a monograph titled Statistical Astrology: The Use of Irregular Seasonality in Quantitative Time Series Analysis and Forecasting. I am currently preparing the research to submit it to an academic/scientific publisher for peer review and publication.
The body of the research is more than 100 pages, and there are more than 350 pages of appendixes.
I've posted the abstract of the monograph below.
This study explores the application of irregular, astrology-based seasonality in quantitative time series analysis and forecasting. The overall objective is to address the inherent limitations of the current approach to seasonality. The noise, the random, irregular component of a decomposed time series, is not nearly as random as it appears, and identifying the patterns and cycles in this noise can potentially improve the overall accuracy of time series forecasting.
The fundamental premise of this study is that seasonality is a quality of time, not of data. Because seasonality is a quality of time, a season can be defined as a discrete unit of time that exists and reoccurs as a subset of a larger recurring unit of time. The larger unit of time is usually linked to the cycle of a reference planet through the zodiac.
The current understanding of seasonality in statistics is based on various divisions of the calendar year, which is the orbit of the Earth around the Sun (i.e., the apparent cycle of the Sun through the zodiac). Months, weeks, days, and hours are all subdivisions of the calendar year, and can all be considered seasons. The uniformity and consistency of these subdivisions of the year constitutes regular seasonality.
When the cycles of planets other than the Earth are used to measure time, the seasonal divisions appear irregular, because to understand the alternative time scales, they must be translated back into calendar and clock units, and the larger cycle that contains the seasons does not correspond with a solar calendar year. This is further complicated by the fact that when viewed from the vantage point of the Earth, as planets in astrology are, the planets appear to slow down and change direction as a regular part of their cycles. The sequence of retrograde periods is its own cycle, independent of the cycle of the planet through the zodiac.
The first part of the study proposes that seasonal influences can be quantified both in terms of effect size, using Cohen’s d and also in terms of the variance of that effect across historical iterations of the season. Regular, Calendar Month seasonality is used as the control, and it is compared with eight other seasonal models across three extensive data sets: Transportation On Time Performance, Car Crashes, and Financial Market Data. The first two data sets carry an expectation of a Calendar Month seasonal influence (the greatest percentage of flight delays occur in December; the greatest number of fatal highway car crashes occur during June and July), while there are no expectations of a Calendar Month seasonal influence in the financial data. Across all three categories of data, the irregular seasonal models had the highest percentage of significant (Medium or greater) effect sizes overall, and when compared to the Calendar Month seasonality. This represents compelling evidence of the presence and potential significance of irregular seasonal influences.
The second part of this study explores the application of irregular seasonality in quantitative time series forecasting. It considers a total of 430 financial forecasts, including 10 stock market indexes, 379 individual stocks, 21 commodities, 10 interest rates and bonds, and 10 currency exchange rates. The forecast period for each data set is 2000–2018. Each forecast is run with twelve different forecast methods, including five standard statistical models (ARIMA, ESM, Holt, Mean, Naïve), a single Mercury-based forecast signal (M15_E3), and six hybrid forecasts, where the seasonal influence of the Mercury forecast is combined with the results of the standard statistical models. The forecasts are run quarterly for the entire 19-year period, and the overall accuracy of each forecast is measured with both MAPE and RMSE. Out of the 12 forecasts for each of the 430 data sets, one of the forecasts that includes the Mercury-based seasonal data was ranked #1 53.49% of the time for MAPE and 45.81% of the time for RMSE, and either #1 or #2 67.67% of the time for MAPE and 62.79% of the time for RMSE. Additionally, the seasonal forecasts were more accurate than their non-seasonal counterparts 71.30% of the time (MAPE) and 66.93% of the time (RMSE), demonstrating that the Mercury-based seasonal influences have significant value in quantitative time series forecasting.
The forecast study is then repeated using a different Mercury-based forecast signal (M15_M3). Out of the 12 forecasts for each of the 430 data sets, one of the forecasts that includes the Mercury-based seasonal data was ranked #1 60.23% of the time for MAPE and 56.28% of the time for RMSE, and either #1 or #2 74.65% of the time for MAPE and 70% of the time for RMSE. The M15_M3 seasonal forecasts were more accurate than their non-seasonal counterparts 74.37% of the time (MAPE) and 72.33% of the time (RMSE).
The final part of this study tests the variance theory of seasonality by re-evaluating the forecasts but only considering the accuracy of the seasons with the lowest variance (and the highest expected predictive value) for each data set. Considering only the targeted seasonality, one of the seasonal forecasts was ranked #1 in accuracy 64.19% of the time for E3 MAPE and 65.12% of the time for E3 RMSE, and ranked either #1 or #2 in accuracy 79.07% of the time for E3 MAPE and 79.30% of the time for E3 RMSE. The targeted seasonal forecasts were more accurate than their targeted non-seasonal counterparts 71.30% of the time (E3 MAPE) and 77.21% of the time (E3 RMSE). One of the targeted seasonal forecasts was ranked #1 in accuracy 65.12% of the time for M3 MAPE and 66.05% of the time for M3 RMSE, and ranked either #1 or #2 in accuracy 77.91% of the time for M3 MAPE and 80% of the time for M3 RMSE. The targeted seasonal forecasts were more accurate than their targeted non-seasonal counterparts 77.53% of the time (M3 MAPE) and 77.07% of the time (M3 RMSE). This provides compelling evidence to support the validity of a variance-based approach to evaluating seasonality.
But the most important conclusion based on the results of this study is that the Mercury-based irregular seasonal model has clear and consistent value in quantitative time series forecasting.