📈 economics

Neural ARFIMA model for forecasting BRIC exchange rates with long memory

This paper proposes a Neural ARFIMA (NARFIMA) model that integrates long-memory structures, nonlinear neural network learning, and exogenous economic drivers to achieve superior forecasting accuracy for BRIC exchange rates compared to conventional benchmark methods.

Original authors: Tanujit Chakraborty, Donia Besher, Madhurima Panja, Shovon Sengupta

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tanujit Chakraborty, Donia Besher, Madhurima Panja, Shovon Sengupta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the Unpredictable

Imagine trying to predict the weather in a city where the climate is chaotic. Sometimes it rains because of a storm system from far away; other times, it's just a sudden, random shower. Now, imagine that city is the global economy, and the "weather" is the exchange rate (the price of one country's money compared to another).

This paper focuses on the BRIC nations (Brazil, Russia, India, and China). These are huge, fast-growing economies, but their money markets are like a stormy ocean: they have long memory (past events affect the future for a long time) and nonlinearity (small changes can cause huge, unpredictable waves).

Traditional tools used by economists are like a simple umbrella. They work okay for light rain, but they fail when the storm gets complex. The authors built a new tool called NARFIMA (Neural ARFIMA) that acts like a high-tech, self-adjusting weather station.

The Problem: Why Old Tools Fail

The paper argues that standard models have two main blind spots:

They forget the long-term: Exchange rates often have a "long memory." If the price was high 10 years ago, it might still influence the price today. Old models often treat the past as if it's irrelevant after a few days.
They miss the hidden patterns: The relationship between oil prices, interest rates, and currency isn't a straight line; it's a tangled knot. Simple math models can't untangle that knot.

The Solution: The "Two-Step" Recipe (NARFIMA)

The authors created a hybrid model that combines the best of two worlds. Think of it as a two-step cooking process:

Step 1: The Linear Base (The ARFIMA Part)
First, they use a statistical model (ARFIMA) to handle the "long memory."

Analogy: Imagine a chef who is great at following a classic recipe. This chef knows that if you add salt today, the soup will taste salty tomorrow and the day after. This step captures the steady, predictable trends and the influence of external factors like oil prices, interest rates, and economic uncertainty (fear in the market).
Result: This step produces a "best guess" prediction, but it leaves behind some "leftovers" (residuals) because it can't explain the weird, chaotic parts of the data.

Step 2: The Neural Network (The "Brain" Part)
Next, they feed those "leftovers" into a Neural Network (a type of AI).

Analogy: Now, a second chef—a master improviser with a super-brain—looks at what the first chef missed. This chef is excellent at spotting weird patterns, like "When oil prices drop and the US is worried about inflation, the currency jumps." This step captures the nonlinear chaos that the first chef couldn't handle.
The "Skip Connection": The model also has a special feature called a "skip connection." Think of this as a direct phone line between the first chef and the final plate. It ensures that the simple, steady trends aren't lost while the second chef is busy fixing the complex parts.

The Ingredients: What Drives the Price?

The model doesn't just look at the currency's past price. It also looks at specific "drivers" that push the price around, much like ingredients in a soup:

Global Economic Policy Uncertainty (GEPU): How worried are people about government rules?
US Market Volatility: How shaky is the US stock market?
Oil Prices: Since Russia and Brazil export oil, and China and India import it, oil prices act like a giant lever on their currencies.
Interest Rate Differences: If a country's banks pay higher interest than the US, money flows in, changing the currency value.

The Results: Who Won the Race?

The authors tested their new "NARFIMA" recipe against 16 other methods, including simple "guess the next number" models, complex statistical formulas, and modern deep learning AI.

The Verdict: NARFIMA won almost every time.
Short-term: It was the clear winner for predicting the next 1 to 3 months.
Long-term: It stayed strong even for predictions 2 to 4 years out, whereas other models fell apart.
The Exception: For India, the data was surprisingly straight and simple (like a calm river), so a simpler model actually worked slightly better. But for the other three countries (Brazil, Russia, China), NARFIMA was the champion.

Why This Matters for Policymakers

The paper explains that Central Banks (the organizations that manage a country's money) need to know where the currency is going to:

Decide how much to charge for loans (interest rates).
Manage the country's savings (reserves).
Prepare for economic storms.

Because the BRIC nations are moving away from relying solely on the US Dollar (a process called de-dollarization), the rules of the game are changing. The NARFIMA model is valuable because it can handle the messy, uncertain, and long-term nature of these new economic relationships better than any other tool currently available.

A Note on Safety (The "Umbrella" Check)

Finally, the paper doesn't just give a single number prediction (e.g., "The price will be $1.50"). It also provides a confidence interval (a range).

Analogy: Instead of saying "It will rain at 2 PM," the model says, "It will likely rain between 1:45 PM and 2:15 PM." This helps policymakers understand the risk and prepare for the worst-case scenario, not just the average one.

In summary: The paper introduces a smart, two-part machine that combines the discipline of traditional statistics with the pattern-recognition power of AI. It successfully predicts the complex, stormy exchange rates of major emerging economies by remembering the long past and understanding the chaotic present.

Technical Summary: Neural ARFIMA Model for Forecasting BRIC Exchange Rates

Problem Statement
Accurate forecasting of exchange rates for emerging economies, specifically the BRIC nations (Brazil, Russia, India, and China), remains a persistent challenge. These series exhibit complex characteristics, including long memory (persistence), nonlinearity, and sensitivity to external macroeconomic drivers. Conventional time series models, such as linear ARIMA or standard ARFIMA, often fail to capture these nonlinear dependencies, while pure deep learning approaches frequently struggle with the theoretical requirements of modeling long memory and non-stationarity, particularly when data availability is limited. Furthermore, exchange rate dynamics in these economies are heavily influenced by global uncertainty measures (e.g., Economic Policy Uncertainty, Equity Market Volatility), oil price fluctuations, and interest rate differentials, necessitating a framework that can jointly accommodate exogenous variables, long-range dependence, and nonlinear interactions.

Methodology
The paper proposes the Neural AutoRegressive Fractionally Integrated Moving Average (NARFIMA) model, an ensemble framework designed to integrate the strengths of fractional integration and neural networks. The methodology follows a two-stage sequential approach:

Linear Long-Memory Modeling (ARFIMAx):
- A linear ARFIMA model with exogenous variables (ARFIMAx) is first fitted to the exchange rate series.
- This stage utilizes historical observations and macroeconomic drivers (Global Economic Policy Uncertainty [GEPU], US Equity Market Volatility [EMV], US Monetary Policy Uncertainty [MPU], oil price growth rates, and short-term interest rate differentials) to capture linear dependencies and long-memory effects via the fractional differencing parameter $d$ .
- The model produces in-sample residuals ( $e_t$ ), which represent the unexplained variations containing nonlinear structures and high-frequency fluctuations.
Nonlinear Residual Modeling (Neural Network):
- A single-hidden-layer feed-forward neural network (ARNN) is trained to model the residuals from the ARFIMAx stage.
- The network inputs include lagged exchange rate observations, lagged ARFIMAx residuals, and lagged values of the exogenous macroeconomic variables.
- The architecture supports two variants: one with skip connections (allowing direct linear influence from inputs to the output) and one without. The skip connection variant is designed to preserve linear dynamics while the hidden layer learns nonlinear interactions, also serving to mitigate vanishing gradients and overfitting.
- Multi-step forecasts are generated recursively.

Theoretical Contributions
The paper establishes the asymptotic stationarity and geometric ergodicity of the NARFIMA process. By modeling the system as a Markov chain and utilizing Lyapunov functions, the authors prove that under specific conditions—bounded activation functions, contractive skip connection weights, and weakly stationary residuals—the process converges to a unique invariant distribution at a geometric rate. These theoretical properties ensure the model's stability and reliability for long-horizon forecasting, addressing a common gap in neural network-based time series literature.

Experimental Setup and Results

Data: Monthly spot exchange rates for BRIC nations (Jan 1997 – Oct 2023) and various macroeconomic covariates.
Causal Analysis: Nonlinear Granger causality tests identified GEPU, US EMV, US MPU, oil price growth, and short-term interest rate differentials as significant drivers of BRIC exchange rates.
Baselines: The model was compared against 16 benchmarks, including statistical models (ARIMAx, ARFIMAx, ETS, GARCH, BSTSx) and deep learning architectures (DeepAR, NBeatsx, NHiTSx, DLinearx, etc.).
Performance Metrics: Evaluation used MAPE, SMAPE, MAE, MASE, and RMSE across six forecast horizons (1, 3, 6, 12, 24, and 48 months).

Key Findings:

Superior Accuracy: NARFIMA consistently outperformed all baseline models across most forecasting tasks and horizons, particularly for Brazil and China. It demonstrated the ability to capture both long-term dependencies and nonlinear patterns effectively.
Robustness: Statistical significance tests, including the Model Confidence Set (MCS) and Multiple Comparisons with the Best (MCB) tests, confirmed NARFIMA's dominance. It remained in the optimal model set across all significance levels for all BRIC countries.
Uncertainty Quantification: The framework was integrated with conformal prediction to generate reliable prediction intervals, offering a model-agnostic approach to quantify forecast uncertainty without strong distributional assumptions.
Residual Importance: Sensitivity analysis confirmed that using ARFIMAx residuals (rather than residuals from simpler models like ARIMAx or Naïve) as feedback to the neural network is critical for the model's superior performance.
Limitations: The model's performance was slightly less dominant for India, where the exchange rate series exhibited more linear characteristics better captured by traditional ARIMAx or linear deep learning models.

Significance and Claims
The paper claims that NARFIMA offers a robust solution for forecasting complex financial time series by bridging the gap between econometric theory (long memory) and machine learning (nonlinearity). Its significance lies in:

Theoretical Rigor: Providing formal proofs of stationarity and ergodicity, which are essential for policy applications where model stability is paramount.
Practical Utility: Demonstrating that incorporating specific uncertainty measures and macroeconomic drivers significantly improves forecast accuracy for emerging markets.
Policy Relevance: Offering central banks and financial institutions a tool capable of navigating the complexities of de-dollarization, geopolitical shocks, and oil price volatility. The authors suggest the model is particularly valuable for managing foreign exchange reserves and formulating monetary policy in volatile environments.
Generalizability: While focused on BRIC exchange rates, the authors posit the framework's applicability to other domains characterized by long memory and nonlinearity, such as interest rates, commodity prices, and epidemiology.

The study concludes that contrary to earlier skepticism regarding the predictability of exchange rates, appropriately structured models that account for long memory, nonlinearity, and external drivers can yield significant forecasting gains.