The Model of Temporal Inertia (MTI) is the decomposition principle that grounds the {4,5,6} Forecast construction. It explains why a value in a time series can be split into two components that operate independently along the sequential timeline and the seasonal timeline: an atemporal absolute (the inertial trend) and a temporal relative (the seasonal relative). The two components can then be combined through Orthogonal Projection combines the two components, and their product is a temporal absolute: a Dimension 4 value at a Dimension 6 calendar coordinate, restoring the {4,5,6} structure of the historical observations. The Microscope for Time is the proprietary library of seasonal models that makes the Dimension 6 component structurally engageable. The seasonal models function as lenses that reveal seasonal structures too complex and irregular for calendar partitions to represent. Temporal Structural Forecasting (TSF) combines the Model of Temporal Inertia and the Microscope for Time; the {4,5,6} Forecast is the result.
The Model of Temporal Inertia
Newton’s First Law of Motion, the Law of Inertia, states that an object at rest remains at rest, and an object in motion remains in motion in constant speed and in a straight line unless acted on by an unbalanced force. This would seem to describe the behavior of time series forecasts where we assume that a pattern of behavior in time series data will continue, except we assume that the Law of Inertia exclusively applies to the dimension of form and not to the dimension of time. We assume this because of the limits of human perception, and specifically because of the limited ability of human beings to perceive the dimension of time.
We accept that the Law of Inertia governs everything about the position of an object in space, but this is not entirely accurate. The Law of Inertia determines the position of an object in space and time. The Law of Inertia already applies to the dimension of time, and therefore, it can function as the foundational theory to explain why time series forecasting works.
The Law of Inertia states that the velocity (speed and direction) of an object moving through space and time remains constant until acted on by an unbalanced force. The Model of Temporal Inertia proposes that the values of data organized in a time series will follow the same trend (speed and direction) until acted on by an unbalanced force.
To understand how time series forecasting works, we require a model. A model makes it possible to connect a hypothesis (the what of an observed event) to a theory (the why).
The velocity (speed and direction) of a moving object can be modeled as an x-t graph (Figure 1), where the y-axis represents the velocity, and the x-axis is time.
Figure 1: Velocity of a Moving Object as an x-t Graph
While the V in Figure 1 might correspond to the velocity of an object in motion, what it actually represents is “value.” The x-t graph that models the velocity of an object in motion is identical in every respect to a plot of time series data. This means that every mathematical formula that defines, explains, and predicts the values of the velocity of an object in motion also defines, explains, and predicts every value in a set of time series data.
The velocity of an object in motion at a single point in time (Vt) is the product of the inertia of that object at that point in time (It) and the cumulative effect of the unbalanced forces acting on that object at that point in time (Ft). We can take this to mean that the value of an observed datum in a time series at a single point in time (Vt) is the product of the inertia of the time series (It) and the cumulative effect of the unbalanced forces (Ft) acting on the time series. We can express this as Vt = It * Ft.
When considering the movement of a physical object, each of these three factors can be observed and quantified independently at each point in time.
We can quantify the speed and direction (velocity) of the object through direct observation. The three main unbalanced forces that act on a physical object in motion are gravity, lift, and drag. As long as we know the physical properties of the object, we can calculate the effects of each of these forces at each moment in time. The inertia of the object depends on the mass of the object, which is constant, and the distance from the axis of rotation, which varies with motion. While it’s possible to calculate the precise inertia of an object, it’s extremely difficult.
Thanks to the Law of Inertia, we don't need to calculate the value of It.
The Law of Inertia states that the value will remain constant over time unless acted on by an unbalanced force. So the inertia of a point in time is equal to the value at the previous point in time. We can express this as It = Vt-1. This means that we can eliminate one factor from the equation and work with Vt = Vt-1 * Ft. The formula to forecast the next value in the series is therefore Vt+1 = Vt * Ft+1. In theory, we can continue with this approach to forecast additional values in the series, using the derivative approach illustrated in Figure 2. The problem with the derivative approach is that the margin of error of each subsequent forecast value increases exponentially.
Figure 2: Individual Relative Values of Ft (Derivative Approach)
The first forecast period always has the lowest possible error.
We can be entirely certain that the inertia of the forecast period is equal to the observed value of the previous period. We can’t know the value of Ft+1, so that’s the factor that introduces the error, Ft+1(E). We can express the forecast like this: Vt * Ft+1(E) = Vt+1(E). The formula to forecast the next value in the series is Vt+1(E) * Ft+2(E) = Vt+2(E2). Because both factors include an error, the margin of error in the second forecast is the square of the margin of error in the first forecast. The error in the third forecast is the cube of the error in the first forecast: Vt+2(E2) * Ft+3(E) = Vt+3(E3). The expected error of each forecast period increases exponentially, and the confidence in each subsequent forecast value plummets accordingly.
If we take an integral approach to forecasting, we can address the problem of exponentially increasing errors.
If we take an integral approach to forecasting, we can address the problem of exponentially increasing errors. The relative value of Ft, which can be expressed as rFt, can be calculated relative to a common denominator, such as the mean value of the time series (MVts) (see Figure 3).
Figure 3: Relative Values of Ft Across Time Series (Integral Approach)
The MVts is a constant, rather than a variable, so it doesn’t introduce a new error into the forecast. The only variables that introduce the errors are the forecasted values of rFt. Each value of rFt has its own margin of error, but those errors are not necessarily dependent on each other and they don’t necessarily increase. The formula to forecast the first value in the series is MVts * rFt+1(E1) = Vt+1(E1). The formula to forecast the second value in the series is MVts * rFt+2(E2) = Vt+2(E2). And the formula to forecast the third value in the series is MVts * rFt+3(E3) = Vt+3(E3). This is possible only if each rFt value is forecast independently and is the product of a single-period forecast. We can’t do this with a single-timeline model, but we can do this if we consider the series of rFt values along the seasonal timeline.
The MVts is an atemporal absolute: it carries the absolute mean value of the time series but no temporal coordinate. Each rFt is a temporal relative: it is indexed to a specific seasonal position and expressed as a ratio rather than an absolute value. The product MVts × rFt is a temporal absolute, recovering the same multidimensional structure as the historical observations. This is the structural mechanism of the Orthogonal Projection: each component carries the quality the other lacks, and the product restores both.
Seasons of Time and the Seasonal Timeline
Human beings model time as a single dimension: a timeline. Your direct experience of time is limited to the present moment. The only correct answer to the question, “What time is it?” is “Now.” It’s always “now.” You remember other nows that are not the now that you experience now, and you call these “then,” and you anticipate nows that you have yet to experience, and you call these “later.” The calendar and the clock provide objective references to organize the sequence of subjective “thens” and “laters,” but your actual, subjective, personal experience and perception of time is limited to the present moment, and the present moment orients you on the sequential timeline. “Now” is a single point on the sequential timeline, but a point has no dimension and therefore can’t be measured. Each unit of measurement of time describes a segment of the sequential timeline that contains an infinite number of points. Each point in time has a corresponding value, which means that the value associated with a unit of time is the mean value of the infinite number of points contained within the segment.
We require a minimum of three coordinates to orient ourselves along the sequential timeline: a period, a season, and a unit.
Each of these represents a segment of the sequential timeline, arranged in a hierarchy. A period is a sub-division of the sequential timeline itself; a season is a sub-division of a period; and a unit is a sub-division of a season. Each segment defines the boundaries of a mean value. The Seasonal Mean Value (SMV) is the mean of the Unit Mean Values (UMV) within a season. The Forecast Period Mean Value (FPMV) is the mean of the UMV within the forecast period. And the Seasonal Relative (SR), the relative effects of the unbalanced forces that we’ve been calling rFt, is the ratio of the SMV to the FPMV (see Figure 4).
Figure 4: Relative Divisions of Time Along the Sequential Timeline
Seasons allow us to orient units within the forecast period.
A season describes one or more of units (points) in time grouped together by a defining characteristic. The seasonal model establishes the defining characteristics that will be used to segment the larger population of time.
When we think of time, we think of the sequential timeline.
The sequential timeline is what is represented by the x-axis when we model time series data. The sequential timeline organizes contiguous events in a linear sequence moving from the past to the future. The sequential timeline is the Dimension 4 axis of the historical record. The seasonal model divides the sequential timeline into contiguous seasons. On the sequential timeline, Season 8 2025 follows Season 7 2025, which follows Season 6 2025.
The seasonal timeline organizes successive, non-contiguous instances of individual seasons within a seasonal model.
The seasonal timeline is the Dimension 6 axis of the historical record. On the seasonal timeline, Season 8 2025 follows Season 8 2024, which follows Season 8 2023.
We’re most familiar with seasonal models that correspond with calendar- or clock-based units of measurement, such as a calendar month or the week of the year. However, the defining characteristics of a season are not limited to the units of the calendar or the clock, and seasons may not be comprised of contiguous and consecutive units (points).
Figure 5 divides the sequential timeline into eight seasons. Each season contains three Unit Mean Values. The SMV of each season is the mean of the three values in the season, and the SR is the ratio of the SMV to the FPMV. To forecast the SMV of Season 5 2025, we need only determine the forecasted Seasonal Relative (FSR) of Season 5 and multiply it by the FPMV.
Figure 5: Seasonal Relatives Along the Sequential Timeline
If we forecast the FSR along the sequential timeline, we're using a derivative approach.
The FSR of Season 5 2025 is dependent on the SR of Season 4 2025. We can be confident of this forecast because it has the smallest possible margin of error. However, we can’t expand the forecast horizon because the FSR of Season 6 would be dependent on the FSR of Season 5, which increases the error exponentially.
If we forecast the FSR along the seasonal timeline, we're using an integral approach.
The FSR of Season 5 2025 is dependent on the SR of Season 5 of 2024. We can forecast the FSR of Season 6 2025 with the same level of confidence because the FSR of Season 6 2025 is dependent on the SR of Season 6 2024. To fully appreciate this, we need a three-axis graph that includes both the sequential timeline and the seasonal timeline (Figure 6).
Figure 6: Modeling Time Series Data with Three Axes
The forecast of each season along the sequential timeline is the product of two single-period forecast values from two different time series on two different timelines.
The single-period forecast for the sequential timeline is the FPMV. The FSR of each season is the product of its own single-period forecast along the seasonal timeline. Because each forecast value is independent, the errors do not increase exponentially and we can have a consistent level of confidence in the forecast for each season along the sequential timeline. This forecast model captures both the trend (mean values within each season) and variability (different mean values between seasons).
Non-Stationary Data and the Model of Temporal Inertia
UU Operations perform best when applied to stationary data, where the values remain relatively constant and exhibit limited variability. When looking for patterns in the historical data, UU Operations must filter out the noise — the so-called random component of the time series data. AI and machine learning variants look for additional patterns in the noise in the hopes of producing a more accurate Prediction, but even so, the more random the time series data is, the less confident we can be in the results of UU Operations.
The Model of Temporal Inertia explains the source of this problem and overcomes it.
Let’s begin by considering what UU Operations actually produce. Non-seasonal UU Operations produce a single Prediction that extends across the entire forecast horizon. This unchanging trend is in fact, the FPMV. In other words, non-seasonal UU Operations are predicting the inertial trend, and ignoring the relative effects of the unbalanced forces.
The FPMV is an atemporal absolute. UU Operations can only produce atemporal absolutes. The relative effects of the unbalanced forces, modeled as seasonal relatives, are temporal relatives: they carry the temporal coordinate at each seasonal position that the UU Operation structurally cannot engage. A UU Operation ignores them not through oversight but through structural incapacity.
To predict the inertial trend, UU Operations must isolate the effects of the unbalanced forces by transforming the data to make it stationary. When viewed through the Model of Temporal Inertia, the non-stationary components of time series data, including trends, cycles, seasonality, and the “random” noise, are the cumulative relative effects of unbalanced forces, which are modeled as seasonal relatives.
In the Model of Temporal Inertia, each forecast value is the product of the FPMV (the inertial trend) and the FSR. It’s a two-step process that requires two forecast values. UU Operations address only step one, the FPMV. The Model of Temporal Inertia addresses the FSR and completes the second step. It doesn’t require stationary data because it’s able to quantify the relative effects of the unbalanced forces and identify WHEN those forces are expected to change (i.e., between seasons).
Seasonal Models: Lenses in the Microscope for Time
Patterns in time exist, but we can’t perceive them directly. Science relies on observation. We can’t understand something if we can’t see it. Sometimes we need new tools to help us to see things that are invisible to the naked eye.
Think of seasonal models as lenses in a Microscope for Time.
When you view a drop of water through the lens of a microscope, you can see a world of single-cell organisms that are otherwise invisible. When you view time series data through the lens of the complex, irregular seasonal models developed by TSF Inc., you can see patterns and cycles that are otherwise invisible. Those patterns allow us to see further into the future with greater detail, precision, and confidence than possible with any existing tool. Each seasonal model is a lens in the Microscope for Time, revealing patterns of unbalanced force along the seasonal timeline.
The ways that we measure and divide time are based on repeated cycles or seasons. A season describes a number of units of time grouped together by defining characteristics. The seasonal model establishes the defining characteristics that will be used to segment the larger population of time. Some of the defining characteristics include calendar-based, contiguous, consecutive, and consistent.
A seasonal model is calendar-based if the seasons are defined as divisions of the calendar year (days, weeks, months, quarters). Every season occurs during every calendar year. If a seasonal model is not calendar-based, a calendar year may not include instances of every season.
A seasonal model is contiguous if the seasons consist of multiple contiguous data points (i.e., a contiguous group of days). A seasonal model is non-contiguous if a season consists of multiple non-contiguous data points (i.e., Mondays in January).
A seasonal model is consecutive if the seasons always follow a fixed order (i.e., the days of the week, the months of the year). A seasonal model is non-consecutive if the order of the seasons can vary.
A seasonal model is consistent if the duration of each season is fixed, unchanging, and equal. A seasonal model is inconsistent if the duration of a season can vary from instance to instance, or if the seasons are not at least approximately equal in duration.
When we think of seasons, we think of the calendar, and it’s hard to think of seasons that aren’t based on the calendar. We have five calendar-based seasonal models to choose from when working with daily aggregated data: Calendar Month, Week Year, Week Day, Month Day (e.g., “Mondays in January”), and Date Month (e.g., 1st of the month, 12th of the month). Each of these is a lens in the Microscope for Time, and each reveals a different set of patterns of seasonal relatives along the seasonal timeline. These are valid seasonal models, but they’re not the only seasonal models.
We can use literally any criteria to define the seasons of a seasonal model, so long as the criteria are objective, the model contains a fixed number of unique, discrete seasons, and each season recurs based on an objective, definable cycle. The more seasonal models we have to choose from, the more patterns we can identify; the more patterns we can identify, the more confident we can feel about the forecasts.
The possibilities are endless. Just as compound microscopes use combinations of lenses to focus, we can use multiple seasonal models to isolate complex patterns in time series data and generate forecasts with greater accuracy and precision.
How Seasonal Models Reveal Temporal Structure
Forecast values with the Model of Temporal Inertia are the product of two single-period UU Operations along two different timelines. The base forecast is the inertial trend, computed along the sequential timeline; it is an atemporal absolute, carrying an absolute value but no temporal coordinate. The seasonal relatives of each season are computed along the Dimension 6 seasonal timeline; each is a temporal relative, indexed to a specific seasonal position and expressed as a ratio rather than an absolute value. The product of an atemporal absolute and a temporal relative is a temporal absolute, recovering the same multidimensional structure as the historical observations. Seasonal models divide the sequential timeline into seasons to determine the seasonal relatives, and provide a matrix to track the historical instances of each season along the seasonal timeline.
The forecast period for the current library of seasonal models is a calendar month. This is the defining container used to compute the seasonal relatives for each individual season. The SMV of each season is the mean of each unit value of each season within the calendar month. The FPMV is the mean value of each unit value of the entire calendar month.
Each forecast value is the product of the base forecast and the Forecasted Seasonal Relative (FSR) of that season. The FSR is the result of a simple moving average of the most recent historical SR values of the season.
Each seasonal model is expressed as a matrix. Each instance of each season has a unique identifier, and this identifier is how each current season is linked with the previous instances of that season along the seasonal timeline.
Let’s consider how this works with the Week Day seasonal model (WD). The WD seasonal model is calendar-based. It consists of seven seasons: the seven days of the week.
The WD seasonal model is non-contiguous. Each season includes either four or five non-contiguous values within each calendar month. The SMV of “Monday” is the average of the daily value of each Monday in that calendar month.
Each specific seasonal instance is designated as [Season]-[Year]-[Quarter]-[Month]. For example, Mon-2024-Q1-Feb, or Thu-2022-Q3-Sep. These unique season identifiers facilitate linking the historical instances of each season across three different series designated S (season), SQ (season-quarter), and SQM (season-quarter-month).
The historical cycles of the WD-S series take the three most recent instances of the season: p1 (one period back), p2 (two periods back) and p3 (three periods back). If the current season is Fri-2026-Q1-Feb, p1 would be Fri-2026-Q1-Jan, p2 would be Fri-2025-Q4-Dec, and p3 would be Fri-2025-Q4-Nov.
The historical cycles of the WD-SQ series take the three most recent instances of the season in the same quarter. So if the current season is Fri-2026-Q1-Feb, p1 would be Fri-2026-Q1-Jan, p2 would be Fri-2025-Q1-Mar, and p3 would be Fri-2025-Q1-Feb.
The historical cycles of the WD-SQM series take the three most recent instances of the season in the same quarter and the same month. If the current season is Fri-2026-Q1-Feb, p1 would be Fri-2025-Q1-Feb, p2 would be Fri-2024-Q1-Feb, and p3 would be Fri-2023-Q1-Feb.
Each of these lenses uses the same seasonal model, but each reveals different historical patterns. Moreover, the amount of historical data becomes a critical factor when selecting which lens to use. The WD-S series has 12 historical instances of each season each calendar year. With 2 years of historical data, every season has at least 24 accuracy metrics, which is statistically significant without becoming unwieldy. Forecasts with the WD-SQ series would each have 6 historical instances with 2 years of historical data, which still clear the minimum requirement of 5 historical instances. But forecasts with the WD-SQM series would each have only 2 historical instances with 2 years of historical data.
With 20 years of historical data, forecasts with the WD-S series would each have 240 historical examples and the forecasted margin of error would be the mean of 240 historical errors. Forecasts with the WD-SQ series would each have 60 historical instances, and forecasts with the WD-SQM would each have 20 historical instances.
A Universe of Seasonal Models
Developing new seasonal models that aren’t based on the calendar is more challenging than you might think. Time is circular and cyclical. To measure time we require some kind of external, objective reference to mark the start of each cycle. Even the calendar and the clock require an external, objective reference. The external, objective references human beings use to measure time are celestial.
The calendar and the clock are based on the observed cycles of the Sun as it appears to orbit the Earth. While days and years are based on the observed cycles of the Sun, the concept of a month is based on the observed cycles of the Moon. We measure and understand time in terms of the calendar and the clock, and we take these systems for granted without considering how and why they were established.
The calendar exists so that farmers can accurately predict the changes of the seasons. The reason for the complicated system of Leap Years in the Gregorian Calendar is so that the Spring Equinox—an external, observable, celestial event—always falls between March 20th and March 21st.
Standardized clock times and time zones exist so that trains can run on time. Until November 18, 1883, when North America adopted “Railroad Time,” which divided the continent into four time zones, each set one hour apart, all time in the United States was local and the country had over 300 individual time zones. And you were today years old when you learned that the branch of the government in charge of time zones and Daylight Savings is the U.S. Department of Transportation.
The calendar-based seasonal models, including the Calendar Month, Week Year, Week Day, and Month Day models, based on the observed cycles of the Sun, are examples of regular seasonality.
Regular seasonality is regular because the duration of the individual seasons of a seasonal model is uniform, the sequence of seasons is consistent, and the cycles of the season fit within the larger container of a calendar year, so that each season occurs at least once during each calendar year.
Seasonal models based on the observed cycles of planets other than the Sun are examples of irregular seasonality. Irregular seasonality is irregular because the duration of the individual seasons of a seasonal model can vary from instance to instance, the sequence of the seasons is inconsistent and variable, and the cycles of the season do not fit within the larger container of a calendar year, so not every season occurs every year.
The other 78 seasonal models in the TSF Inc. library are complex, irregular seasonal models based on the observed cycles of the Moon, Mercury, and Venus. These lenses include models with as many as 204 and as few as 3 seasons. The average duration of individual seasons ranges from 1 to 30 days with most seasons consisting of from 4 to 15 contiguous calendar days.
The {4,5,6} Forecast
The Model of Temporal Inertia decomposes a time series value into the FPMV (the inertial trend) and the FSR (the relative effect of the unbalanced forces). The FPMV is an atemporal absolute: it carries an absolute value but no temporal coordinate. The FSR is a temporal relative: it preserves the temporal coordinate at each seasonal position and expresses the value as a ratio rather than an absolute level. The two components forecast independently along structurally distinct timelines: the FPMV along the sequential timeline, the FSR along the Dimension 6 seasonal timeline. Independent single-period forecasts maintain consistent confidence across the forecast horizon.
The Microscope for Time supplies the seasonal partitions that make Dimension 6 structurally engageable. At each forward position, the construction applies the lens with the strongest prior track record at the corresponding seasonal position, determined mechanically from the historical record. The lens library is fixed; nothing about the construction is optimized to historical outcomes.
Orthogonal Projection combines the two single-period forecasts into a {4,5,6} Forecast. The UU Operation along the sequential timeline produced the FPMV (an atemporal absolute); the UU Operation along the Dimension 6 seasonal timeline produced the FSR (a temporal relative). Their product is a temporal absolute: a Dimension 4 Forecast Value located at a Dimension 6 calendar coordinate, surrounded by a Dimension 5 Calibrated Probability Band (CPB). The Forecast Value recovers the {4,5,6} structure of the historical observations because each component supplies the quality the other lacks: the FSR supplies the temporal coordinate that the FPMV stripped; the FPMV supplies the Dimension 4 value that the FSR normalized away. No amount of scaling, sophistication, or AI applied to a single UU Operation changes the categorical distinction: a Prediction is an atemporal absolute; only the Orthogonal Projection of two operations produces a temporal absolute.
The CPB is computed as the empirical distribution of forecast-actual differences at the corresponding seasonal position across prior cycles, with the current position excluded from the calibration. The 85% band contains 85% of those historical differences; the 95% band contains 95%. The label and the coverage rate match because the band is built from the values themselves.
Every component of the {4,5,6} Forecast is empirical, derived from the historical record, and structurally locked against retroactive adjustment.