Masterclass: Forex Through the Lens of Science
Trend (or trend component) — is a somewhat conventional concept. A trend refers to the systematic, non-random component of a time series—typically monotonic—that can be computed via a well-defined, unambiguous rule. Trend analysis and duration estimation are central to a trader’s workflow. Therefore, mathematically modeling the probability of trend existence, monotonicity, and likely continuation is a highly relevant task.
Trends in time series often reflect underlying physical laws or other objective regularities. Yet it is impossible to definitively separate a random process or time series into a regular (trend) component and an oscillatory (residual) component. Consequently, the trend is usually assumed to be a simple functional form—linear, quadratic, etc.—that captures the ‘overall behavior’ of the series. If such a simplification aids analysis, the chosen functional form is considered acceptable.
Fig. 1. Trend on the weekly USD/RUB chart.
After extracting a linear trend, its statistical significance must be assessed—typically using correlation coefficient analysis. A correlation coefficient significantly different from zero suggests a real (positive or negative) trend—but this difference may arise purely by chance due to sampling variability in the observed segment. In other words, analyzing another segment of the same time series could yield a correlation estimate much closer to zero—or even of opposite sign—making the existence of a true trend ambiguous.
Specialized statistical techniques exist to test trend significance. One widely used approach applies the Student’s t-distribution.
Fig. 2. Trend on the weekly EUR/USD chart.
Beyond linear trends, more complex trend structures frequently appear. Pre-specifying an appropriate functional form is rarely feasible. In practice, analysts often test several simple parametric functions (e.g., linear, exponential, quadratic), evaluating how well each fits the observed trend. With modern computing, these evaluations are fast—and sometimes fully automated—to identify the optimal functional form among candidates. However, none of the tested functions may adequately capture the underlying trend. In such cases, alternative approaches are required: transforming the series (e.g., logarithmic transformation, differencing—computing successive differences, or integration—cumulative summation) to induce a clearly linear trend. Once achieved, standard linear trend methods apply; results are then back-transformed to the original scale. Crucially, the trend is not uniquely determined by the data—it is a conditional construct introduced to better understand the process.
It is challenging to automate and model trend detection in time series. However, if the trend is monotonic (consistently rising or falling), analysis becomes tractable. When measurement noise is substantial, the first step in trend extraction is smoothing—a local averaging process where nonsystematic components cancel out. The most general smoothing method is the moving average, where each data point is replaced by a simple or weighted average of neighboring points within a window of width L. Instead of the mean, the median of values in the window may be used. The key advantage of median smoothing, versus moving average, is robustness to outliers (e.g., measurement errors). Thus, when outliers are present, median smoothing typically yields smoother—or at least more reliable—curves than a moving average with identical window width. Its main drawback: in outlier-free data, median smoothing produces more ‘jagged’ curves and does not support weighting schemes.
When measurement error is exceptionally large, alternative smoothing methods may be applied: least-squares smoothing (weighted by distance), or negative exponentially weighted smoothing. All filter noise and produce relatively smooth curves. Short, regularly spaced series may be smoothed using splines. Nevertheless, many monotonic time series are well approximated by linear functions. If a clear monotonic nonlinear component exists, data should first be transformed to remove nonlinearity—commonly via logarithmic, exponential, or (less frequently) polynomial transformations.
A superficial familiarity with economic theory often breeds contempt for specialized knowledge—a view noted by Nobel laureate M. Friedman. Practically speaking, any forecast based on the expectation that current conditions will persist—whether a chain of setbacks or successes, or a worsening or improving situation—is, in essence, a bet on the law of inertia. Unsurprisingly, this principle has long been observed in financial markets. Market price movement can be modeled as a combination of two states: inertia of rest or inertia of motion—once initiated by an external impulse (macroeconomic shifts, market psychology, randomness, etc.), motion continues beyond the initial impetus. In the 1960s, numerous scientific works mathematically justified trend persistence, aligning with the broad notion of ‘economic inertia’—i.e., the systemic inability of an economy to abruptly reverse course. Analyzing market participant behavior reveals that sentiment cannot instantly shift from bullish to bearish (or vice versa), giving rise to the concept of market inertia. The sheer scale of the market prevents rapid maneuvering; all strategic moves must be initiated well in advance—or they simply fail on the FOREX market. While markets constantly execute ‘turns’, their motion remains predictable thanks to market inertia. Predicting the exact direction of movement at any given moment remains difficult—but assuming a trend exists simplifies forecasting. With reasonable confidence, one can estimate where the exchange rate is likely to be at the next time step. A sharp reversal is possible, but requires extraordinary force—further explaining FOREX market inertia.
FOREX Forecasting Method: SSA
One powerful technique for time series analysis and forecasting is Singular Spectrum Analysis (SSA). SSA serves both as an analytical tool and a filtering method. Below is a concise overview of its operation.
To analyze a time series, a parameter L—called the “window length”—is selected. L may be chosen arbitrarily. For sufficiently long series and large L, results become largely insensitive to window length. The SSA “Caterpillar” method solves multiple tasks: trend extraction, periodicity detection, series smoothing, and full decomposition into trend, periodic, and noise components. This versatility—under minimal assumptions—comes at two costs: first, grouping singular value decomposition components of the trajectory matrix is highly manual and non-automated; second, the absence of an explicit statistical model prevents formal hypothesis testing about component presence (an inherent limitation of all nonparametric methods). For example, linear analysis of a currency pair revealed nonstationarity, ruling out high-quality parametric forecasting models. Applying the nonparametric “Caterpillar”-SSA approach yielded more promising results: the resulting model correctly identified the dominant directional bias. However, it failed to forecast fluctuations within the primary trend. To improve forecast quality, researchers deviated from theoretical recommendations and adjusted parameters intuitively—but this did not enhance performance. Ultimately, the theoretically grounded model was confirmed as optimal—with its associated forecast accuracy limit.
Accordingly, representing the trend of a fundamental indicator as a symbolic string composed of fuzzy variables introduces a novel element that simplifies formal description of trend characteristics and enables further processing—e.g., using structural pattern recognition methods, proven effective in technical diagnostics for assessing interactions between macroeconomic dynamics and exchange rates across multiple fundamental indicators FI(xi), where xi ∈ X. Such assessments allow analysts to isolate only those causal relationships among FI(xi) where one indicator demonstrably influences the trend of another—effectively compressing the conceptual map of the macroeconomic system under study.
Given that market participants commonly assess influence between fundamental indicators by comparing trend similarity [6, 7], the Euclidean distance between quantitatively defined trends—a standard procedure in pattern recognition theory—is a natural choice for measuring such similarity.
When trends are described qualitatively—as symbolic strings of fuzzy variables—the appropriate tool is linguistic and syntactic analysis of experimental curves. This involves constructing a symbol dictionary (representing possible trend states) and a grammar (defining valid combinations of symbols to describe curve shapes). Building grammatical parse trees for symbolic trend strings—using a formal grammar—enables a novel, simple, and well-formalized approach: evaluating similarity between parse trees to recognize influence of one fundamental indicator on another’s trend. Structural similarity/dissimilarity measures for complex systems follow specific construction rules and must be applied accordingly.
Literature
FAQ
What is the main challenge in mathematically defining a trend?
A trend is not uniquely determined by the data—it is a conditional, model-dependent construct, often approximated by simple functional forms (e.g., linear or exponential) because real-world time series cannot be definitively separated into trend and noise components.
Why is median smoothing preferred over moving average in some cases?
Median smoothing is more robust to outliers (e.g., measurement errors), yielding more reliable trend estimates when data contamination is present—though it produces less smooth curves than moving averages in clean data and does not support weighting schemes.
How does Singular Spectrum Analysis (SSA) help in FOREX trend analysis?
SSA decomposes time series into trend, periodic, and noise components without assuming a parametric model, enabling nonstationary trend extraction and directional bias identification—though component grouping is manual and formal statistical testing is not supported.
