Modeling the probability that a currency exchange rate will reach a specified target level over a future time horizon is framed as an extrapolation problem. This approach leverages the advantages of mathematical modeling for multi-stage information processing, solution method development, and transformation into executable programs.
Stages and Problem Formulation in Mathematical Modeling
Mathematical modeling is rarely linear. It’s iterative: revisiting earlier stages with adjusted inputs to progressively refine the model. Stage one assesses the real-world situation using available prior knowledge and the modeling objective. Stage two produces a conceptual model—expressed in domain-specific language (economics, physics, etc.)—that captures the problem’s essence. Stage three selects the model structure: the appropriate mathematical framework, number and form of equations, and functional forms. Stage four refines the model—applying approximations and calibrating coefficients. Stage five evaluates model quality using criteria tied directly to the modeling goal. If performance falls short, the process loops back—to the start or an intermediate stage—for another iteration.
Despite infinite variation in objects, contexts, and goals, these core modeling stages recur. Work begins with all available information: experimental data (from the object or similar systems), established theory, expert intuition. It proceeds through preliminary analysis of observed time series—and ends with applying the model to solve a concrete task. Because models evolve, this is almost always iterative: repeated returns to earlier steps, successive refinements toward a robust result.
Two qualitatively distinct cases arise from how the time series is generated. First: observations are numerical realizations of a known mathematical model—e.g., solutions to differential equations. Here, “reconstruction” or “recovery” of governing equations is meaningful. Validation is straightforward: compare against the true underlying equation. Theoretical conditions for method effectiveness can be derived. Second: the time series originates from measurements of a real-world process. In this case, no single ‘correct’ model exists, and modeling success cannot be guaranteed.
Extrapolation as the Initial Stage of Final Forecast Construction
One can only marvel at the “unreasonable effectiveness of mathematics” if a high-quality model emerges [2,3]. For example, in mid-1990, a team of American students visited Las Vegas’s most powerful casino—and returned home with several million dollars. Over weekends for one month, these sharp engineering students played blackjack and won enormous sums. They were from the world-renowned Massachusetts Institute of Technology (MIT). Their million-dollar casino wins reignited a dormant war between players and casinos—a conflict originally launched 40 years earlier by mathematics professor Edward Thorp. Thorp realized blackjack was the most mathematically tractable casino game: unlike roulette, dice, or lotteries—where past outcomes do not influence future ones—blackjack exhibits memory and dependency due to card depletion.
A fundamental challenge in gambling is identifying betting opportunities with positive expected value. An analogous problem in investing is finding investment opportunities offering “excess” return—i.e., risk-adjusted returns above a benchmark. Once such favorable opportunities are identified, both gamblers and investors must decide what fraction of capital to allocate. Interest in “excess return” dates back at least to the 18th century and Daniel Bernoulli’s discussion of the St. Petersburg Paradox. Yet money management remains equally critical. On stock markets (including securities markets), the problem is similar—but more complex. The gambler becomes the investor, seeking “high returns at a controlled risk level.” In both cases, Edward Thorp applied the Kelly Criterion—which maximizes the expected logarithm of wealth (“expected logarithmic utility”) [4].
The general extrapolation problem consists of estimating values of a function describing how a variable evolves over time—beyond the interval where the function has been observed—enabling its use for forecasting. Today, extrapolation is widely employed as a method for generating forecasts from simple predictive models. It projects future trends under the assumption that the patterns observed historically will persist. These patterns define the most stable feature of the forecasted process—the trend—assumed to be describable by some functional form.
Fig. 1. Trading indicators as signals of future trends on Forex.
In applied mathematics, a typical extrapolation task is as follows: given the first N terms of a sequence {z}, can we:
1) Determine whether the sequence converges;
2) Estimate its limit;
3) Quantify the error in that estimate;
4) Assess the reliability of the error estimate using a formal criterion?
At first glance, this appears unsolvable—too little information is available. Indeed, the first N terms alone do not uniquely determine the full sequence; beyond term N, behavior may be arbitrary. Such problems are termed ill-posed in mathematics. Yet in practice, they arise constantly: predicting the position of a moving body, forecasting system operating conditions, and many design and control tasks fall into this category. Examples show that the initial terms of a sequence often contain far more information about its limit than commonly assumed—the key is extracting it effectively.
One common error estimation technique compares a computed value with its extrapolated counterpart. Extrapolation is also used to accelerate convergence of sequences. Well-known acceleration methods construct a new sequence from the original {z} that converges to the same limit—but faster. In some cases, acceleration yields results unattainable within reasonable time by other means. Statistical forecasting methods hold a central place among prediction techniques. Their application assumes that the pattern driving historical development (within the time series) will continue into the forecast horizon—i.e., forecasting relies on extrapolation. Forecast accuracy depends critically on how valid the assumption is that the factors shaping the series’ core components in the base period will remain operative in the future. Therefore, any extrapolation-based forecast must be preceded by careful analysis of long-term time series to reliably identify the underlying trend. Since trends themselves may evolve, extrapolated results must be treated as probabilistic estimates—not deterministic certainties.
Extrapolation into the future is called prospective; into the past, retrospective. When referring to time-series extrapolation, prospective extrapolation is usually implied. Its use in forecasting rests on two key assumptions: (1) overall development follows a smooth curve; and (2) the dominant historical and current trend will not undergo major shifts going forward. Thus, forecast reliability and precision hinge entirely on how realistic these assumptions are—and how accurately the historical pattern has been characterized. Extrapolation should therefore be viewed as the foundational, initial stage in building final forecasts.
Using Time Series for Statistical Forecasting Methods
Solving any problem via statistical modeling entails:
- Developing and constructing a structural process diagram, identifying key interrelationships;
- Formally describing the process;
- Modeling stochastic phenomena accompanying the system—random events, variables, or functions;
- Accumulating simulation results, performing statistical processing, analysis, and generalization.
The results of statistical modeling are subject to experimental error. Consequently, any conclusion about the modeled system must be supported by appropriate statistical tests. Experimental error in statistical modeling depends heavily on the fidelity of stochastic phenomenon modeling. As known, probabilistic systems involve random events, random variables, and random functions—so modeling stochastic phenomena reduces to modeling those three elements. Since random events and random functions can be expressed via random variables, their simulation ultimately relies on generating appropriate random variables. Hence, a mathematical model of trader behavior is required—one incorporating memory of past decisions and capable of selecting actions based on win-probability estimates. This aligns with the foundational hypothesis of collective behavior theory: “any sufficiently complex behavior arises from combinations of simple behavioral acts.” Basic deterministic dependency-recovery methods start from a given time series—i.e., a function defined at a finite set of time points. A vast literature addresses time-series analysis and forecasting. Time series are often embedded in probabilistic models that include additional factors (independent variables)—e.g., money supply (M2 aggregate). A time series may be multivariate, with multiple outputs (dependent variables). Core tasks are interpolation and extrapolation (i.e., forecasting). The least-squares method—in its simplest form (linear function of one variable)—was developed by German mathematician C. Gauss in 1794–1795. Preliminary variable transformations may improve performance. For financial market participants, this approach is known as technical analysis.
Statistical forecasting methods require long time series. Thus, they are impractical in rapidly changing environments or when forecasting newly emerging situations. As an alternative, expert forecasting methods—relying on specialists’ experience and intuition—are used, especially adapted to the mathematical modeling problem of exchange rate dynamicsexchange rates
FAQ
What is the core challenge in modeling the probability of reaching a target exchange rate?
The core challenge is treating it as an ill-posed extrapolation problem: limited historical data cannot uniquely determine future behavior, requiring assumptions about trend persistence and careful error estimation.
Why is extrapolation considered only the initial stage—not the final forecast?
Because extrapolation assumes historical patterns will continue unchanged; real-world exchange rates are influenced by evolving fundamentals, shocks, and structural shifts, so extrapolated results must be probabilistic and refined with additional modeling or expert judgment.
How does the Kelly Criterion relate to exchange rate modeling in this context?
It is cited as an analogous framework for optimal capital allocation—just as it guides bet sizing in blackjack based on win probability, it informs risk-adjusted position sizing when trading based on probabilistic exchange rate targets.
