Integrating AI tools and VAR models for actuarial applications: an analysis of EGP/USD exchange rate dynamics Open Access

The formalization of currency exchange systems began in the nineteenth century with fixed exchange rates and the gold standard, later evolving through the Bretton Woods Agreement of 1944, which institutionalized fixed rates. This system collapsed in the early 1970s, leading to the adoption of floating exchange rates. Exchange rates are commonly classified into nominal and real types: the nominal exchange rate refers to the market rate of exchange between two currencies, while the real exchange rate reflects the relative price of domestic versus foreign goods (Roy, 2021). MacDonald (2007) further distinguishes nominal rates into spot and forward rates. The spot exchange rate (S) denotes the price for immediate currency delivery, typically within one to two days, whereas the forward exchange rate (F) is the pre-agreed rate at which a currency will be exchanged at a future date, generally around 90 days, and negotiated between involved parties.

Currency exchange is subject to continued transformation, shaped by technological innovation and global financial dynamics. Future contracts, as noted by MacDonald (2007), extend the concept of forward nominal exchange rates by allowing present-day agreements for future currency delivery. Simultaneously, the emergence of digital currencies like Bitcoin is already altering foreign exchange markets, with their influence expected to grow. As globalization prevails, exchange rate management significantly affects currency strength and macroeconomic stability, influencing key indicators such as GDP, trade, inflation, interest rates and financial markets. Consequently, exchange rate volatility holds substantial economic implications, prompting central banks and policymakers to closely monitor these fluctuations to maintain stability.

The dynamic interplay between exchange rates, inflation and interest rates remains a central focus in economic and actuarial research, given its relevance to macroeconomic stability and accurate projections (Laing, 2008; Louberge, 1983; Chang and Lee, 2020; Chang et al., 2019).

In Egypt, recent volatility in the exchange rate between the Egyptian Pound (EGP) and the US Dollar (USD) has significantly influenced inflation and interest rates, underscoring the need to closely monitor these relationships. Such volatility is especially pertinent for actuaries, as fluctuations directly affect projected cash flows in investment and insurance operations. Understanding these variables’ interdependencies enables more precise actuarial risk assessments and valuations. For example: in insurance companies operating in Egypt, foreign exchange exposure—arising from reinsurance treaties, USD-denominated premiums and claims and foreign currency assets—necessitates incorporating exchange rate analysis into actuarial assumptions. Accurate estimations of inflation and interest rates are also essential for discounting future liabilities and asset valuation. Therefore, considering exchange rate behavior in actuarial modeling enhances the reliability of cash flows projections and supports effective risk management.

This paper presents a comprehensive analysis of the Egyptian Pound (EGP) versus the US Dollar (USD) exchange rate over a critical period in the Egyptian economy, from early 2017 to mid-2023. The primary objective is to offer actuaries enhanced insights into exchange rate trends for improved projection accuracy, while also exploring the causal relationships between exchange rates, inflation and interest rates within the broader macroeconomic context. To achieve this, a novel hybrid methodology combining quantitative and qualitative approaches is employed including:

1. Time-series modeling and causality tests applied to structured quantitative data to identify statistical relationships.

2. State-of-the-art Artificial Intelligence (AI) tools, particularly Natural Language Processing (NLP) techniques, to extract insights from textual unstructured data within financial narratives and reports that contextualize economic trends.

To support the analysis, two distinct datasets covering January 2017 to July 2023 are utilized. The first comprises structured quantitative data, including monthly EGP/USD exchange rates, inflation and interest rates for both Egypt and the United States, sourced from the Central Bank of Egypt (CBE) and the Federal Reserve. The second consists of unstructured textual data drawn from monthly reports issued by the Egyptian Ministry of Finance, offering qualitative insights into the socioeconomic and political factors influencing Egypt’s economy, especially in relation to inflation and interest rate dynamics.

By adopting this hybrid analytical approach, the study offers a novel approach combining classical quantitative approaches with qualitative approaches based on state-of-the-art AI tools. This approach provides a comprehensive assessment of the trends and interdependencies among exchange rates, inflation and interest rates, enhancing actuarial understanding for projection and risk evaluation. The integration of advanced AI tools complements traditional quantitative techniques by incorporating qualitative insights from economic reports, thereby equipping actuaries with a timely and multidimensional strategy for interpreting the broader financial and macroeconomic context.

The paper is structured as follows: Section 2 offers a thorough overview of previous studies on currency exchange rate behavior and its economic implications, on one hand, and on exchange rate risk and its impact on actuarial applications, on the other hand. Section 3 presents the statistical modeling and empirical results. Section 4 introduces the AI methodologies employed and summarizes the qualitative findings; and Section 5 concludes with final remarks and directions for future research.

The dynamics of currency exchange rates and their influence on economic growth have drawn considerable interest from academics, governments and policymakers. Exchange rate volatility—driven by policy decisions, geopolitical developments and market speculation—reflects the fluctuation in currency values over time, prompting fiscal and monetary interventions to maintain stability. Hausmann et al. (2004) note that real exchange rates in developing economies are nearly three times more volatile than those in advanced economies, a disparity not fully explained by larger economic shocks or crisis susceptibility. Hassan and Zhang (2020) further underscore the significance of exchange rate stabilization policies and trade deficit sustainability in shaping cross-border capital flows. This discussion, though selective, offers a focused synthesis of the relationship between exchange rate volatility and macroeconomic performance, particularly in the context of developing economies.

In Egypt, the Egyptian Pound (EGP) is closely tied to the United States Dollar (USD), making the EGP/USD exchange rate and its stability critical for the Egyptian economy. This relationship is reflected in GDP calculations, where GDP at current prices is converted to USD using the yearly average of market exchange rates.

Data from the World Bank and the International Monetary Fund (IMF) indicate that Egypt’s GDP rose from approximately 5 billion USD in 1965 to 78.78 billion USD in 2004, followed by a sharp increase to 393.91 billion USD in 2023. Projections suggest continued growth, with GDP expected to reach 539.61 billion USD by 2029, marking a 55.24% increase over five years. These trends emphasize the pivotal role of exchange rate stability in sustaining economic expansion and enhancing productivity. A recent study by Dalgic and Ozhan (2024) in Turkey further reinforces this link, revealing a strong correlation between GDP and exchange rate movements, as well as between inflation and exchange rate fluctuations. The study concludes that in recessionary contexts, exchange rate depreciation often results in an appreciation of dollar-denominated assets, serving as a hedge against economic contraction.

With the increasing integration of global markets and the expansion of international trade networks, the influence of exchange rate fluctuations on trade balance has become increasingly pronounced. Currency appreciation typically reduces the cost of imports while making exports less competitive, potentially impacting trade performance. In 2022, Egypt’s exports of goods and services represented 10.56% of GDP, whereas imports accounted for 19.30%. In January 2024, the Egyptian government reported a 68% reduction in the trade deficit, indicating a positive trend in export growth and a decline in imports (Ministry of Trade and Industry, 2024), a development expected to contribute to greater EGP/USD exchange rate stability.

Comparable patterns are observed in other economies; for instance, the Central Bank of Exchange Rate Management in Nigeria (2021) reported that currency devaluation typically raises the cost of imports, leading to inflationary pressures and potentially slowing economic growth.

The interaction between inflation, interest rates and exchange rate volatility is also discussed in the literature. Governments often intervene in currency markets to curb inflation and stabilize interest rates to mitigate adverse macroeconomic impacts. In Pakistan, Ali et al. (2015) found both short- and long-term correlations between inflation and exchange rate movements, noting that increased inflation often drives higher interest rates and expanded money supply, which in turn amplifies exchange rate volatility. Similarly, Asari et al. (2011) examined data from Malaysia (1999–2009) and concluded that interest rate hikes effectively reduce exchange rate volatility, highlighting the intricate relationships among these key economic variables.

Roy (2021) emphasizes that flexible exchange rate regimes enable governments to implement expansionary monetary policies to counteract recessions; however, this flexibility also generates uncertainty in international transactions, potentially undermining economic stability. In contrast, fixed exchange rate systems restrict monetary policy autonomy, necessitating coordinated efforts among governments to sustain stability.

Recognizing the importance of these dynamics, our study prioritizes modeling exchange rate trends in the Egyptian market and their interactions with inflation and interest rates, as elaborated in Section 3. These considerations highlight the critical need to analyze exchange rate trends to promote stability and mitigate volatility risks especially in insurance companies or investment banks. These institutions, either multinational or domestic, operating in emerging markets, the fluctuations of exchange rate can substantially affect performance and profitability.

Therefore, financial and insurance experts, including actuaries, must rigorously assess exchange rate behavior and its causal relationships with inflation and interest rates to offer strategic guidance to governments and market stakeholders. The critical role of currency risk management is especially evident in actuarial applications within the insurance sector, which is discussed in the subsequent section.

Across global markets, the volatility in currency exchange rates creates the currency risk. Foreign currency risk exposure is a critical consideration in financial and actuarial practice. According to Collier and Davis (1985), currency risk can be classified into three categories:

1. Transaction Risk – The risk of gains or losses arising from financial transactions involving currency conversion.

2. Translation Risk – The risk associated with the revaluation of foreign investments in financial statements.

3. Economic Risk – The broader impact of currency volatility on future cash flows from foreign investments.

Despite the acknowledged importance of forecasting exchange rate movements to manage currency risks and volatility, Maurer and Valiani (2007) underscore the difficulties of predicting exchange rate fluctuations, particularly in economies with flexible exchange rate regimes. The impact of exchange rate dynamics on the insurance industry in Egypt and other African markets, with an emphasis on actuarial practice, is outlined herein.

Foreign currency risk for insurance companies in developing markets has been studied highlighting the critical role of exchange rate trends in mitigating risk exposure and maintaining financial stability. In the Egyptian context, Nada and Ibrahim (2017) analyze foreign exchange exposure among insurance firms, examining its effects on key performance metrics such as profitability, loss and investment ratios from 2000 to 2016. Additionally, Amer (2014) emphasizes that a significant portion of Egypt’s reinsurance transactions are denominated in foreign currencies, directly affecting actuarial calculations, including incurred but not reported (IBNR) reserves. Furthermore, Ismail and Sharkawi (2024) explore the ramifications of exchange rate liberalization on long-term premiums, claims settlement, investment returns and the overall functioning of the Egyptian insurance sector.

Moreover, several studies in African insurance markets examine the long-term impact of exchange rate volatility. Offiong et al. (2020) analyze Nigeria’s insurance industry, recommending a unified exchange rate regime to mitigate currency fluctuation risks. Ipigansi and Egoro (2022) assess how exchange rate changes affect claim settlements in Nigeria from 1990 to 2018, while Kemuma (2015) explores the influence of exchange rate movements on the profitability of Kenyan insurance firms.

In contemporary financial settings, many life, non-life insurers and pension funds operate in multiple currencies and maintain diversified portfolios including foreign assets. Laing (2008) notes that such exposure introduces additional risks beyond domestic investments, impacting cash flow valuation, asset-liability management, solvency and profitability. Moreover, the Society of Actuaries (SOA) in its Speciality Guide “Actuarial Aspects of Currency Risk (1997)” emphasizes the importance of actuaries evaluating exchange rate fluctuations alongside inflation and interest rates, highlighting the critical interplay of these variables for actuarial computations in financial and insurance institutions. Actuaries’ expertise in financial risk assessment and modeling positions them as key actors in managing currency risks within insurance companies, pension funds and financial institutions, with significant applications related to currency risk management. Key actuarial applications related to currency risk include:

1. Mathematical and Statistical Modeling:

   • Developing actuarial projections using statistical models and based on historical data and macroeconomic indicators.

   • Using assumption-based models incorporating interest rates and inflation trends (Khalil, 2014, 2015) to assess surplus/deficit scenarios in actuarial valuations.

2. Monitoring and Reporting:

   • Establishing continuous monitoring frameworks for currency exchange rates and their impact on the expected cash flows including investment returns.

   • Analyzing interest and inflation rate trends to enhance economic assumptions for risk assessment.

3. Hedging Strategies:

   • Designing currency hedging strategies to mitigate potential losses arising from exchange rate fluctuations.

   • Assessing an organization’s risk tolerance and providing strategic recommendations to policymakers in insurance and financial institutions.

Actuaries also play a crucial role in ensuring compliance with financial regulations concerning currency exchange, requiring them to remain informed about monetary policies, international financial standards and regulatory frameworks to effectively manage currency risk. Additionally, a thorough understanding of government policies affecting exchange rate movements and their wider implications for financial decision-making is essential.

Here, it is worth mentioning that the devaluation of domestic currencies is clear evidence of the impact of currency exchange risk in actuarial work. The case of Egypt is a good example of that. Upon the devaluation of EGP against USD since 2016, the Egyptian Central Bank has decided to raise the interest rate. This is also accompanied by increasing dramatically the inflation rate. The government was trying to combat inflation and the devaluation of domestic currency. The decision to increase the interest rate was mainly to contain the devaluation of the domestic currency and to attract foreign currencies for domestic investments to boost the national economy as a whole. The impact of currency risk on actuarial work in the Egyptian insurance market affects two main activities:

1. projections of cash flows in insurance operations and

2. investment returns.

To mitigate the currency risk associated with these activities, actuaries are advised to adopt the following:

1. Ensuring the implementation of statistical modeling for projections, such as: time-series analysis of currency exchange rate trends, to understand and estimate the future cashflows in the operations conducted using foreign currencies. This could help in getting more accurate calculations regarding the reserves’ estimates and profitability.

2. Integrating best estimate assumptions for currency exchange, interest rates and inflation indices into actuarial modeling in the process of preparing the actuarial valuation.

3. Diversifying the investment portfolios and encouraging operating in domestic and other foreign currencies.

4. Encouraging and facilitating good investment channels in foreign currencies which can be supported by regulatory authorities to reinforce the domestic currency.

5. Recommending monitoring the currency exchange rate by the government to control the currency risk.

6. And finally, coping with the socioeconomic environments by conducting a qualitative analysis based on the information in economic and policy reports that can be extracted by AI tools.

Thus, AI tools using NLP techniques would be a helpful tool as they are capable of extracting relevant information from these reports. This would assist actuaries in interpreting government policies, thereby broadening the comprehension of economic dynamics related to currency fluctuations that affect the insurance companies. By integrating these advanced AI methods with traditional statistical techniques such as vector autoregression (VAR) models and Granger causality tests—detailed in the following section—this research aims to deliver a comprehensive hybrid analysis that enhances the precision and effectiveness of actuarial evaluations concerning currency exchange risk.

Over the past few decades, the modeling of currency exchange rates using statistical techniques has been a crucial area of study for economists, financial experts and actuaries. Understanding currency exchange rate trends, their volatility and the impacts on economies has gained substantial significance, especially when considering the interdependencies between exchange rates, inflation and interest rates. Various methodologies have been employed to understand the fluctuations in currency exchange rates, estimate future expectations and manage associated risks. This section explores the theoretical foundations and techniques commonly used for modeling exchange rates and their application to understanding currency exchange volatility.

Historically, one of the simplest methods for understanding currency exchange rate trends and volatility is through the definition used by option traders, which characterizes volatility as the standard deviation of daily fluctuations in exchange rates. Hagen and Zhou (2007) emphasize that estimating volatility in historical exchange rate data typically involves modeling the standard deviation or variance of past exchange rate movements to project future volatility. This approach provides an essential measure of past exchange rate variability, which can be crucial for projecting future behavior.

However, the volatility of currency exchange rates presents significant challenges, particularly in projecting future fluctuations. Researchers have long struggled with making accurate predictions due to the inherent nonlinearity in exchange rate behaviors. Early studies by Meese and Rogoff (1983), Boothe and Glassman (1987) and Wolff (1987, 1988) have shown that traditional linear econometric models fail to outperform simple random walk models in terms of projecting accuracy. As a result, the use of nonlinear techniques has become increasingly prevalent, with numerous studies (e.g. Hsieh, 1989; Papel and Sayers, 1990; Diebold and Pauly, 1988; Meese and Rose, 1991; Chinn, 1991) suggesting that nonlinear models offer superior forecasting outcomes.

Several time series models have been widely adopted for analyzing exchange rate dynamics, notably the ARIMA and ARCH family of models. ARIMA models are particularly effective for capturing patterns in univariate time series and forecasting future values based on historical trends and autocorrelations. However, their univariate nature constrains their ability to incorporate interactions among multiple economic variables simultaneously. Similarly, ARCH models and their generalized extensions, such as GARCH (Engel, 1982; Bollerslev, 1986; Bollerslev et al., 1992), are extensively employed to model conditional heteroscedasticity and volatility clustering frequently observed in financial time series. Despite their strength in modeling time-varying volatility, GARCH models also typically operate within a univariate framework and primarily emphasize volatility dynamics rather than the structural interrelationships among macroeconomic variables. Consequently, both ARIMA and GARCH models exhibit limitations in capturing the intricate interdependencies among exchange rates, interest rates and inflation relationships that are crucial for comprehensive macroeconomic and actuarial analyses.

In the realm of stochastic modeling, techniques such as Brownian motion and geometric Brownian motion are often employed to simulate exchange rate movements under the assumption that currency exchange rates follow a random process. These models are widely used to generate potential scenarios for future fluctuations based on the historical randomness of the exchange rate.

While these models provide valuable tools for understanding exchange rate dynamics, another prominent approach involves conditioning the models with economically meaningful variables. Incorporating macroeconomic factors such as interest rates, inflation rates, GDP growth and geopolitical events is essential for improving the predictive accuracy of exchange rate models. However, empirical studies in this area remain varied, with some research suggesting the difficulties in accurately predicting exchange rates, especially in the short term (Alagidede and Ibrahim, 2016).

Incorporating macroeconomic variables in the models is critical for long-term exchange rate predictions. For instance, studies by Mark and Sul (2001) indicate that exchange rates exhibit a long-term positive relationship with macroeconomic fundamentals, meaning that these variables tend to move in the same direction over extended periods. However, predicting exchange rate movements in the short run remains a complex task (Corte et al., 2009; Li et al., 2015) with research by Li et al. (2015) assessing the out-of-sample exchange rate predictability based on a random walk model by using one-month ahead forecasts generated by recursive regressions.

Several studies have examined the interplay between currency exchange rates and macroeconomic indicators such as interest rates and inflation. These interrelationships may be varied according to period, country and economic environment. For example, Ali et al. (2015) applied the Johansen co-integration test, Granger causality and impulse response function (IRF) to analyze the relationships between exchange rate volatility, inflation, interest rates and money supply in Pakistan. Their study found that exchange rate volatility is influenced by inflation, interest rates and money supply in both the short and long run.

Furthermore, Mohamed (2023) examined the effects of exchange rate and interest rate dynamics on macroeconomic variables such as unemployment, economic growth and inflation in Egypt from 1991 to 2020. The study applied the Engle–Granger two-step co-integration model and found that both the exchange rate and interest rate affected inflation in the long term, while the interest rate had a more direct effect on inflation in the short term.

In summary, a wide range of models and statistical techniques have been used to analyze the trends and interrelationships between currency exchange rates, inflation and interest rates. The choice of model depends on various factors such as data availability, model assumptions, market conditions and the specific needs of the stakeholders. In this study, the vector autoregressive model (VAR) and Granger causality tests are adopted to explore the interrelationships between the exchange rate, inflation and interest rates. The VAR model is adopted in this study, as it treats all variables as endogenous and is well-suited for analyzing the joint behavior and causal relationships among the variables over time. These models will be applied to the Egyptian context to better understand the dynamics at play during the period from 2017 to 2023, with the results presented in the following subsections.

Multivariate time-series data extend univariate time series by modeling multiple time series simultaneously. The primary objectives of multivariate time-series analysis are to explore the dynamic relationships between variables and improve prediction accuracy (Tsay, 2005). One of the most versatile and widely used methods for modeling multivariate time series is the VAR model.

The VAR model, introduced by Sims (1980), is a well-established econometric tool for analyzing the dynamic interactions among multiple time-series variables. A VAR model generalizes the univariate autoregressive model, allowing for the projection of a vector of time series. This model provides a systematic framework for methodically comprehending the inter wind interaction of multivariate time series data, which is why it is extensively applied in various economic and financial analyses. Unlike structural models, which impose a priori theoretical restrictions, VAR models are more flexible for empirical analysis as they do not require such constraints.

The equation for a VAR model of order p, denoted as VAR(p), can be expressed as:

where $Yt=(Y1t,...,YKt)′$ is a (K × 1) random vector, the $Ai$ are fixed (K × K) coefficient matrices, $ν=(ν1,...,νK)′$ is a fixed (K × 1) vector of intercept terms allowing for the possibility of a nonzero mean E(⁠$Yt$⁠) and $ut=(u1t,...,uKt)′$ is a K-dimensional white noise or innovation process (Lütkepohl, 2005).

The VAR model treats all variables symmetrically as endogenous. By estimating $Ai$ we can examine dynamic relationships between the variables. Forecasts are produced from a VAR system recursively. Each variable within the system is forecasted.

Granger causality, proposed by Granger (1969), is a statistical test used to determine whether one time series can predict another. The idea is that a cause cannot come after the effect. Thus, if a variable x affects a variable z, the former should help improve the predictions of the latter variable (Lütkepohl, 2005). It should be noted that Granger causality does not imply true causation; it only identifies predictive relationships based on temporal precedence.

Under the Granger causality test framework, we compare two models. A restricted model (without X) given by:

And an unrestricted model (with X) given by:

where $p$ represents the lag length, $αi$⁠, $βi$ and $γi$ are the estimated coefficients, $ϵt$ and $νt$ are the error terms which are assumed to be white noise.

The null hypothesis is that $Xt$ does not Granger-cause $Yt$ and the alternative hypothesis is that at least one of the $γi≠0$⁠, implying that $Xt$ Granger-cause $Yt$⁠.

Together VAR model and Granger causality provide a robust framework for understanding interdependencies and causal relationships in multivariate time-series data.

This subsection focuses on the statistical modeling based on structured data, which includes the monthly exchange rate series of the Egyptian pound per US dollar (EGP/USD) along with interest and inflation rates covering the period from January 1,2017 to August 1, 2023. The three-month treasury bill rates of Egypt and the United States are used as proxies for interest rate in Egypt and United States, respectively. Table 1 shows the description of the variables included in the study together with their abbreviations.

The inclusion of US data is essential for this study, as the exchange rate being analyzed is the EGP/USD. Exchange rates are inherently comparative, reflecting the macroeconomic conditions of both currencies involved. By incorporating the US datasets, the study ensures a comprehensive evaluation of the relative movements in economic indicators between Egypt and the United States, allowing for a more accurate assessment of their impact on the exchange rate dynamics.

The ER, INTEG and INFEG data are obtained from the Central Bank of Egypt (CBE). As for INTUS and INFUS data are obtained from the Federal Reserve Bank (FED), USA.

Table 2 presents the descriptive statistics for the variables, revealing high variability in the exchange rate (ER) series and the interest rate in Egypt (INTEG). This variability is primarily attributed to the devaluation of the Egyptian pound during the study period.

The time-series plot for ER, INTEG and INTUS is shown in Figure 1. The data are not stationary with an overall upward trend. From January 2017 to January 2022, no clear trend is observed, followed by a sharp increase in the ER attributed to the devaluation of the Egyptian pound during that time. It is also evident that the series exhibit parallel movements, suggesting a long-run relationship between the interest rate and the exchange rate.

To further analyze the evolution of the EGP/USD exchange rate relative to inflation rates, the percentage change in the EGP/USD exchange rate was computed and plotted against the core inflation rate in Egypt (INFEG) and United States (INFUS) as shown in Figure 2. INFEG appears to move in tandem with ER. The high spikes were during the period of the EGP devaluation. The percentage change in ER and INFEG appears to be non-stationary. However, the series for INFUS appear to be stationary.

Since the EGP/USD exchange rate is relative—i.e. it compares the currencies of the two countries—both the interest rate and inflation rate are likewise computed as relative representing the difference between the rates in Egypt and United States as follows:

Figure 3 shows the time-series plot for the percentage change in the exchange rate (ER) along with the relative interest rate (RINT) and relative inflation rate (RINF). The plot illustrates that the variables are not stationary.

Since the VAR model requires a stationary series of data, Augmented Dickey–Fuller (ADF) and Phillips–Perron (PP) tests for stationarity were conducted for all variables. Table 3 presents the results of these tests for the variables before and after first differencing (d = 1). Both ADF and PP test results show that the p-values for ER, RINT and RINF are less than the 1% level of significance after first differencing. Therefore, it can be concluded that the data become stationary after first differencing.

Accordingly, the first difference of the exchange rate, RINT and RINF series are used in constructing the VAR model, as they are all stationary as evidenced by the ADF and PP test.

To select the optimal number of lags to include in the VAR model, the Akaike Information Criterion (AIC), Hannan Quinn (HQ) and Schwarz Criterion (SC) were computed for different numbers of lags. The model with the lowest information criterion value is considered optimal.

Table 4 displays the findings, indicating that AIC suggests the best model should include five lags, whereas both HQ and SC recommend the best model should include two lags. Given that a five-lag model would complicate the analysis, the two-lag model is chosen based on the HQ and SC results. Thus, a VAR (2) model is performed with three estimated equations (one for each variable). These equations are structured as follows:

The VAR equation for the exchange rate (ER) in Table 5 is statistically significant at the 1% level based on the F-statistic (6.309, p < 0.001), with an adjusted R² of 0.293. This indicates that approximately 29.3% of the variation in the EGP/USD exchange rate is jointly explained by its own lags and the lags of the RINT and RINF.

The VAR equations for the RINT and RINF are also statistically significant at the 10% and 5% levels, with F-statistics of 1.968 (p = 0.082) and 2.711 (p = 0.020), and adjusted R² values of 0.070 and 0.118, respectively.

The significance of individual coefficients varies across equations. For instance, the second lag of ER is highly significant in both the ER and RINT equations, while the second lag of RINF is significant in the RINF equation at the 1% level.

The Granger causality tests were performed to examine the causal relationships among ER, RINT and RINF. The results of the tests are presented in Table 6. Each relationship was tested using both the standard F-test and the Chi-square (Wald) test. Accordingly, two p-values are reported: one for the F-statistic (⁠$PF$⁠) and one for the Chi-square test (⁠$Pχ2$⁠).

The results show that the RINT and RINF combined can be used to predict the EGP/USD exchange rate (⁠$Pχ2$ = 0.066). In addition, a unidirectional causality is observed from the EGP/USD exchange rate to RINT (⁠$Pχ2$ = 0.007) implying that the EGP/USD exchange rate can be used to predict RINT. Moreover, both the EGP/USD exchange rate and RINF combined can be used to predict the RINT (⁠$Pχ2$ = 0.039). However, there is no causality detected for RINF.

Examining the system’s IRFs and variance decomposition is a standard procedure in the VAR model. The IRF quantifies how shocks affect endogenous variables over time, either directly or indirectly through another variable. Their primary objective is to explain how a model’s variables change in response to a shock to one or more of the variables. The orthogonal impulse responses for ER, RINT and RINF are provided in Figures 4, 5, and 6. The y-axis represents the percentage variation, while the x-axis represents the time in months. To enhance the interpretability of these results, 95% bootstrap confidence intervals have been included in the IRF plots. These intervals are critical for assessing the statistical significance of the estimated responses. Specifically, an impulse response is considered statistically significant only if its confidence band does not include zero. Highlighting these intervals provides a more robust understanding of which shocks generate meaningful effects and the time horizon over which these effects persist.

Figure 4 plots the estimated response of ER, RINT and RINF respectively to a forecast error impulse in ER with bootstrapped error bands. At the initial period, a positive shock in the exchange rate causes ER to rise by approximately 0.6%. The shock diminishes with the impact returning to almost zero in the 8th month. A one standard deviation shock in the exchange rate has a minimal effect on both RINT and RINF. However, there is a noticeable spike in RINT at the time the exchange rate increases.

Figure 5 shows the estimated response of ER, RINT and RINF respectively to a forecast error impulse in RINT with bootstrapped error bands. In the first period, there is no impact of RINT shock on the ER. The impact is highest in the third period with a negative response of ER to RINF shock. A one standard deviation shock in the RINT has a small effect on RINT and RINF.

Figure 6 shows the estimated response of ER, RINT and RINF respectively to a forecast error impulse in RINF with bootstrapped error bands. There is a positive response from the ER in the second month when there is a shock in RINF which converges shortly after the 4th month. A one standard deviation shock in the RINF has a minimal effect on RINT and RINF.

The quantitative analysis highlights key relationships between the EGP/USD exchange rate, interest rates and inflation. The results indicate that a two-lag VAR model provides the best fit for capturing these interactions. The Granger causality test further reveals that relative interest and inflation rates combined can predict exchange rate movements, while the exchange rate itself influences interest rates.

The immediate and significant response of the exchange rate (ER) to its own shocks likely reflects strong exchange rate persistence and the role of market expectations during periods of devaluation or policy uncertainty in Egypt. The delayed response of ER to shocks in relative inflation (RINF) suggests that inflationary pressures may influence exchange rate movements indirectly, possibly through their impact on purchasing power or investor confidence. In contrast, the relatively weak and delayed response of ER to shocks in RINT may indicate limited monetary policy transmission during the study period, potentially due to capital controls or a lack of investor responsiveness to domestic interest rate changes.

However, statistical models alone do not fully capture the broader economic forces shaping these trends. From an actuarial perspective, while quantitative analysis is the main basis of actuarial analysis and valuation, a more comprehensive understanding of the underlying economic context requires additional and valuable insights. Financial reports, policy statements and expert commentaries/reports contain crucial qualitative information that can enhance and complement the interpretation of statistical findings. In the following section, we turn our attention to the use of AI tools to analyze unstructured financial data, leveraging NLP techniques to extract relevant economic indicators and trends that can be of main interest to actuaries.

Recently, actuaries and financial experts have shown a growing interest in integrating advanced AI tools into their analytical workflows.

So far, our study has focused on a quantitative analysis of currency exchange rate and its relationships with inflation and interest rates using numerical structured data as detailed in Section 3. However, to achieve a comprehensive understanding, we will complement this with a qualitative approach. In this section, our analysis involves leveraging AI-driven tools to analyze unstructured data, offering deeper insights into the trends of exchange rate, and the dynamics of the interrelationships with interest and inflation rates within economic and financial context.

Although actuarial studies are traditionally rooted in structured data analysis, they often require insights from economic reports to complement statistical findings. These reports, typically written in natural language, contain valuable qualitative information that could enrich actuarial decision-making. However, processing these extensive reports manually is often time-consuming and challenging due to:

1. Volume of Information – Economic, financial and actuarial reports are usually lengthy and detailed, which takes more effort to extract relevant insights efficiently.

2. Complexity of Interpretation – Understanding macroeconomic trends and socioeconomic factors requires a nuanced contextual analysis, which can be difficult to achieve solely through structured datasets.

3. Time Constraints – Actuaries and financial analysts often work within tight deadlines, necessitating automated solutions for processing vast amounts of qualitative data.

NLP and AI have revolutionized how unstructured data is processed. Transformer-based Large Language Models (LLMs), such as GPT (Generative Pretrained Transformer), have enabled the automation of tasks such as summarization, classification, translation and information extraction from long documents.

Since the launch of ChatGPT in November 2022, Generative AI has reshaped the way professionals interact with unstructured data through new paradigms. Over the past two years, the AI “arms race” has accelerated, with major technology firms, such as Microsoft, Google and Meta competing to develop more advanced LLMs and AI-driven analytical tools.

The backbone of modern LLMs is the Transformer Architecture first implemented by a Google Research team (Vaswani et al., 2017). These models can perform various NLP tasks, which could be applied to the actuarial domain, such as:

1. Summarizing financial reports.

2. Extracting key insights from economic publications.

3. Analyzing trends in exchange rate policies.

4. Identifying correlations between macroeconomic indicators in text.

In this context, the question raised here is how do actuaries could benefit from these advances in analyzing the trends of currency exchange rate and its interrelationships with inflation and interest rates? Could AI assist actuaries in the projection modeling of currency exchange rates from contextual reports? Finally, to what extent are the AI-generated insights reliable in actuarial contexts?

For this study, a hands-on experiment was conducted aiming at assessing how reliable are the new AI tools in assisting actuaries in their analysis. The experiment is based on a well-known NLP task; “Question Answering” (QA). The main idea behind QA systems is: Given a text corpus and a question in plain language, the system automatically provides the answer.

The objective of the experiment is to assess the effectiveness of AI tools in processing economic reports and extracting meaningful insights relevant to currency exchange rate trends. The findings from this study will provide a data-driven evaluation of how AI can support actuarial decision-making in foreign exchange risk assessment.

Extractive and abstractive approaches are central to question answering (QA) systems. Extractive QA involves retrieving direct text fragments from source documents (Mallick et al., 2023), whereas abstractive QA, particularly with LLMs, generates coherent, human-like summaries from the context, offering broader applicability and more natural responses (Zafar et al., 2024). While domain-specific fine-tuning of LLMs for actuarial and insurance contexts would enhance accuracy, it necessitates extensive training data, computational resources and expert oversight—limitations that preclude this path in the current study. Instead, this research evaluates the usability and reliability of existing AI tools in supporting actuaries in currency exchange rate analysis through an accessible and efficient approach.

To achieve this goal, the study employs zero-shot prompting. To clarify this concept, “prompts” are the instructions or questions introduced to the AI model, and the AI model responds to these prompts. For some tasks, few examples are first given to the AI model to customize and adapt the output to the required task. In zero-shot prompting no examples are given, as it leverages the inference capabilities of pre-trained AI models without requiring retraining or domain-specific data (Kojima et al., 2022). This advantage makes it practical for actuaries by avoiding resource-intensive model customization, allowing immediate application of generative AI tools and facilitating timely, adaptable analysis. Prompts can be interrogative (e.g. “What is the interest rate?”) or imperative (e.g. “Please indicate the reported interest rate”).

For this study, the experiment was carried out using Gemini, the pre-trained generative AI interface provided by Google (2024) as detailed in the experiment section. Gemini is free access, user-friendly interface and reliable. Opting for different AI interfaces is possible, yet it is beyond the scope of this study. For future work, it is foreseen to carry out further experiments using multiple AI interfaces to compare results.

As discussed earlier, understanding exchange rate trends and their interrelationships with inflation and interest rates requires a deep contextual understanding of the economic and political landscape shaping the dynamics between these variables. Traditionally, economic reports contain valuable unstructured data that provide insights into the variables under study. However, manual analysis of such reports is time-consuming – subject to human biases – challenging due to complex financial language. To address these challenges, AI-based text processing tools have demonstrated significant advancements in analyzing large volumes of economic text. Thus, we designed an experiment to evaluate how AI tools could be utilized by actuaries regarding their evaluation of currency exchange rate, interest and inflation rates by leveraging methodologies commonly applied in NLP studies. The next subsection outlines the experimental setup and methodology used in this research.

This experiment follows a structured methodology commonly applied in AI-based research, particularly in NLP studies. The process consists of four key phases explained below.

1. Dataset compilation: Compiling a set of financial reports, as unstructured data sources, that contain relevant information on currency exchange rates, interest and inflation rates within the same time frame set for the study, i.e. reports covering from year 2017 to year 2023. In terms of NLP, the compiled reports represent our corpus which serves as the unstructured textual dataset used as an input for the AI model.

2. Questions and prompt design aligned to actuarial interests: Once the dataset was compiled, we formulated a set of targeted questions designed to extract key insights from the financial reports. These questions were carefully structured to be used within prompts or instructions given to the AI interface to inquire about information related to the three variables under study from the reports reflecting the perspective of an actuary. This is to ensure that the responses generated by the AI tool align with actuarial decision-making needs.

3. Zero-shot prompting: For this experiment, we utilized Gemini 1.5, a generative AI interface based on a LLM developed by Google (formerly known as Bard) and released in 2024 (Google, 2024). Gemini is a pre-trained model with a transformer architecture. It is a generative model for general purpose tasks with a conversational interface with the same name, Gemini. The conversational interface is publicly available for free. Gemini Pro is another paid version offering more capacities for the model. The initial trials were carried out by the freely available Gemini, while the final experiments were conducted with Gemini Pro.

4. Analysis and evaluation: The final step is to qualitatively analyze and evaluate the AI-generated outputs. The evaluation process focused on four key criteria:

   • Accuracy – Does the response correctly extract information from the reports?

   • Relevance – Are the insights meaningful and applicable to actuarial analysis?

   • Coherence – Is the response structured and logically sound?

   • Comparability – How does the AI-generated response compare to traditional actuarial interpretations?

Figure 7 represents the different phases of the experiment and the following subsections will provide a detailed description of each of these phases.

To run the experiment, it is necessary to create a corpus of unstructured data as input for the AI model to process. The compiled reports used in this study are a sample selected from the monthly financial reports published by the Egyptian Ministry of Finance. These reports are made publicly available in English through their official portal. For each year from 2017 to 2023, two reports are selected: December report and June report. The decision to select these two months is driven by their financial significance: The January report provides insights into the economic situation since the beginning of the year, and the June report reflects the state of the economy at the end of the fiscal year in Egypt during the observation period.

These reports are instrumental in assessing economic trends and identifying key financial indicators that affect the currency exchange rates, inflation rates and interest rates, which are of particular interest for actuarial applications.

The overview of the nature of this dataset reveals general characteristics that we summarize as follows:

1. Reports vary in their length; older reports tend to be shorter while the more recent reports are slightly longer. The length of reports is an important aspect as some AI models have input size limitations.

2. The language used in the reports is quite systematic with a tendency of long sentences leading to some ambiguity. This makes it harder to identify and extract relevant information. The use of passive voice and negations can further complicate the task, as it may obscure direct answers or conclusions.

3. All reports include a section for monetary perspective where different key indicators affecting currency exchange rates are usually mentioned within the text.

4. The introductory section of each report usually sums up main contextual information on the socioeconomic and the political conditions affecting the financial and economic environment.

Before designing the prompts for the AI tool, a set of 31 questions were prepared. From the 31 questions, 25 are unique questions while the remaining 6 questions are variations of the unique questions to test different ways of formulating the question and if the answers received would vary by changing the way the question is formulated. These questions are specifically designed to extract information related to the currency exchange rate, interest rate and inflation rate from the compiled financial reports. They are grouped into two main subsets based on the content and purpose:

1. First subset: the actuarial perspective: the questions reflect the types of insights an actuary would be interested in extracting from the reports, i.e. the practical information that actuaries would use for their valuations.

2. Second subset: quantitative analysis: the questions are formulated to be aligned with the statistical analysis, i.e. specifically focusing on trends and cause–effect relationships between the variables under study.

The questions are formulated taking into consideration the following:

1. Question format and complexity: Different ways of formulating the question:

   • Simple Yes/No questions (e.g. Does the inflation rate increase in the following report?)

   • Wh-questions such as What, How, etc. (e.g. What is the currency exchange rate in the following report?)

2. Affirmative instructions (e.g. Explain how does the monetary policy affect the currency exchange rate in the following text)

3. Question type according to the information being asked:

   • Recall/descriptive questions or direct questions whose answers are explicitly stated in the report (e.g. Does the interest rate increase in the following report?)

   • Quantitative questions inquiring about quantitative information, i.e. numeric rates or values (e.g. What is the inflation rate in this report?)

   • Qualitative questions inquiring about qualitative information factors, reasons or broader insights (e.g. What are the factors that affect the currency exchange rate in the following report?)

   • Inference/analytical questions or indirect questions whose answers are not explicitly stated in the report, but are based on making inferences or analyzing relationships from information mentioned in the report (E.g. How does the interest rate react to shocks in EGP/USD exchange rate and inflation rate based on the report provided?).

     1. Questions inquiring about the relationships between the variables under study.

     2. Questions inquiring about trends of the three variables.

     3. Question inquiring about predictions of the currency exchange rates.

     4. Question inquiring about the evaluation of international bodies about the performance of the Egyptian economy.

With the set of questions, a prompt was designed including:

1. The description of the objective of the task.

2. The report in text format.

3. Instructions on how the AI should carry out the task.

4. The set of questions.

5. Instructions on the answer’s format.

As mentioned earlier, zero-shot prompting is adopted for this experiment, i.e. no examples are given to the conversational interface of Gemini. This process is carried out through different iterations covering the monthly reports of the period under study. An example of the prompt is shown in Figures 8 and 9. Figure 8 provides the description of the task and the instructions to follow while Figure 9 includes a sample of the questions asked on each report.

In this section, the answers generated by the AI interface are thoroughly analyzed based on the four evaluation criteria mentioned above to assess the advantages and limitations of using AI tools in assisting actuaries in analyzing financial reports and extracting useful insights.

The main questions we are trying to answer through this evaluation are the following:

1. Advantages of using AI tools (relevance, precision and groundness of the answers)

   • Could AI tools assist actuaries in the arduous task of analyzing reports and extracting the relevant information from large amounts of text?

   • To what extent are the answers provided precise and well-grounded?

2. Limitations of AI tools (reasoning, predictions, inferences)

   • Can the AI tool extract coherent comparisons from the report?

   • Are AI tools able to infer trends over time based on information in reports without being explicitly mentioned?

   • Are AI tools able to make predictions based on information in reports without being explicitly mentioned?

To answer these questions, we carried out an analysis of 425 responses provided by the AI tool (answers to 25 unique questions distributed throughout 17 separate reports). It is a qualitative evaluation where each response is evaluated in terms of its accuracy, relevance, coherence and comparability to what a human actuary would extract from the report. Based on this extensive analysis, we are able to identify the main advantages and limitations of using AI with examples from the evaluated answers.

After in-depth analysis of each question and the evaluation conducted, it is confirmed that AI tools provide a substantial opportunity for actuaries to save time and effort in analyzing reports. Instead of reading through long texts manually, actuaries can rely on AI tools to extract relevant information quickly, which helps reduce the time required to analyze and summarize reports. Examples of correct answers are provided in Table 7. The main advantages are the following:

1. AI tools demonstrate an ability to extract specific pieces of information regarding the rates of the variables under study with accuracy in most cases. For example: in Recall/Descriptive questions or direct questions whose answers are explicitly stated in the report.

2. AI tools tested in this experiment are able to make inferences, reason and compare between different indicators, and hence, draw a full picture of the performance of the economy, although all the information is not explicitly provided in the report. For example: in a cause–effect relationship.

However, the detailed analysis revealed clear limitations that actuaries and professionals in the insurance industry should be aware of to make the appropriate use of the tools and not to take the given answers for granted. Examples of answers considered incorrect or imprecise are provided in Table 8. This is why, it is indispensable to highlight the limitations of these tools:

1. AI tools tested in this experiment are not usually precise in extracting numeric values such as interest rates, overnight deposit rates. This might be due to the fact that in the reports used, these rates are provided in comparison with other periods or other months which makes it more difficult to identify the current rate under study.

2. AI tools are not able to predict future changes in the rates given the information included in the reports. For example, in inference/analytical questions regarding predictions of rates.

3. AI tools are not able to infer trends from the report. For example, in inference/analytical questions regarding inference of trends.

4. In some cases, the answers are repeated. This might be a limitation of the AI tool, but also it might, in some cases, be related to the nature of language used in these monthly reports as it tends to adopt a systematic and repetitive structure.

5. To deal with the errors, actuaries should keep in mind that AI tools are for assistance only and should always be used under human supervision keeping usually the expert in the loop. Errors in the answers cannot be directly corrected except in two scenarios that require more technical knowledge from the actuary to deal with AI interfaces:

   • The actuary in case of error could further iterate with more questions and prompts by indicating explicitly that the answer provided is wrong and by re-formulating the prompt. In this way, the AI tool might adjust its answers. This technique is called Reinforcement Learning from Human Feedback (RLHF). However, this technique should be carried out systematically by technical professionals known as prompt-engineers.

   • The second optimum scenario is to fine tune the existing AI tools for the actuarial task. This is feasible, but as mentioned requires technical, computational resources and should be funded and sponsored by institutions or by funded projects.

Finally, generative AI tools leveraging LLMs offer significant advantages for actuaries in the insurance industry and other financial sectors. They help save considerable time and effort when processing unstructured data such as financial reports, and they offer valuable insights by extracting precise, relevant information. These tools are particularly beneficial for straightforward queries and explaining cause–effect relationships about variables like inflation, exchange rates and interest rates.

However, certain limitations must be taken into account, especially in tasks requiring precise numeric data, deep reasoning, trend analysis and complex predictions. While AI is excellent for summarizing and reasoning from past data, it may not yet be a substitute for expert judgment in predictive modeling.

Thus, while AI tools hold promise for streamlining the actuarial process in terms of speed and efficiency, they should be viewed as tools for assistance, always under human supervision rather than a replacement for expert judgment and domain-specific knowledge.

With vast financial transactions worldwide, analyzing the trends of the currency exchange rates has had a great influential role in the economies. Focusing on the impact of volatility in currency exchange rate on interest and inflation rates is of an eminent importance to governments and decision makers. Thus, modeling the trends of currency exchange rates and interrelationships with interest and inflation rates are of main interest among economists, experts in the insurance industry and actuaries.

Being specialized in modeling risks, actuaries are not only interested in understanding the trends of currency exchange rate, interest and inflation rates for projection purposes of expected cash flows in insurance companies and investment firms, but also in being well-informed of the dynamics of these variables in the financial markets.

In our paper, two approaches are adopted to provide actuaries with a comprehensive analysis of the trends of the three variables under study in Egypt during the period from January 2017 to August 2023. They are: (1) a quantitative statistical approach based on applying the VAR and Granger causality test on structured data and (2) a qualitative approach using the NLP and advanced AI tools for processing unstructured data. Both structured and unstructured data are collected during the study period from different sources. The structured data is obtained from the Central Bank of Egypt and the Federal Reserve Bank – USA. The unstructured data is attained from the monthly reports publicly published by the Ministry of Finance-Egypt.

The results of implementing the VAR model show that the best model can include two lags. The Granger causality test reveals that interest and inflation combined can be used to predict the EGP/USD exchange rate; in addition, a unidirectional causality from the exchange rate to interest rate is observed. Furthermore, both exchange rate and inflation rate combined can be used to predict the interest rate. From a qualitative perspective, the AI advanced interactive interface of Gemini by Google (2024) is used to extract information related to the variables under study from textual contents of the financial reports. Thus, the experiment is conducted based on a well-known NLP task using the “QA” technique and zero-shot prompting. The findings reveal that using AI tools could be helpful to the actuaries in answering Wh and Yes/No questions. In addition, it extracts correct information about the description of the variables which is mentioned explicitly in the text. Moreover, it is able to provide explanations about cause–effect relationships between the variables under study with an overview of the economic environment in Egypt. However, a few limitations of using the AI tool are detected. Most importantly, they are not able to predict future movements precisely or infer trends of the variables.

Amid the high progress in AI and the emergence of more powerful tools such as: “Deepseek”, we can further explore the usage of more AI tools to compare if they are capable of extracting more precise answers. The optimum scenario is to fine-tune a domain-specific AI model and a LLM to the actuarial domain and to a certain task. Nevertheless, as pointed earlier, requires extensive computational resources and a technical team behind. In our opinion, this is an important line for future work that should be considered as a multidisciplinary project which could be funded by the academia or by the industry.

In this respect, it is important to point out the ethical concerns when using AI tools. These tools should always be considered as assistants for humans, not to substitute humans. So, keeping the human expert in the loop is essential. In addition, these tools should be used bearing in mind that they can generate content from other sources, not only from the inputs provided during the experiments. So, generated output should not be considered as ground truth and should always be critically evaluated.

Finally, an important conclusion is that adopting a hybrid approach would highly contribute to decision making at the economic and actuarial levels. Comprehensive analysis considering different data sources and different approaches provides the necessary knowledge for well-informed decisions. This is why the approach presented in this study would also assist policymakers in analyzing the vast amount of information from multiple sources and different formats.