GDP forecasting means predicting the future size of an economy, measured by gross domestic product, the total value of everything a country produces. It matters because major decisions rest on it. Central banks set interest rates partly on where they expect the economy to go, governments plan budgets and spending around it, and investors and businesses size their plans to the outlook. A forecast that is wrong in the wrong direction can be costly, so the method used to produce it matters, and that is what this paper tests. This study reflects MCC Economics’ econometric and data-science work on the Gulf economy, built on a dataset of 426 economic variables from more than 20 sources, and the full paper is published on the Munich Personal RePEc Archive, SSRN and Emirates Scholar.
Table 1: Five-year GDP forecast accuracy by model, measured by mean absolute percentage error where lower is better, tested on United Arab Emirates data from 1975 to 2023.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
You use a statistical model that learns from past data and projects it forward, and there are several families of these models to choose from. This paper compares three of the most widely used and least expensive, so that any organisation could run them. ARIMA, short for autoregressive integrated moving average, looks only at the past pattern of GDP itself and extends it. VAR, short for vector autoregression, tracks GDP alongside other economic measures and how they move together. Linear regression links GDP to a set of driving factors, such as oil production and exports, and estimates the relationship. Each has strengths, and the paper’s job is to find which suits the United Arab Emirates best.
It depends on the horizon, and the paper measures this precisely using a metric called mean absolute percentage error, which is simply the average size of the forecast’s miss expressed as a percentage. Lower is better. Over five-year windows the time-series models won, with ARIMA best at 5.2 percent average error and the stronger VAR next at 7.1 percent. Over one-year windows ARIMA again had the lowest average error at 3.0 percent, but the linear regression model, when fed accurate inputs, was the most consistent, which matters as much as the average for a decision-maker. The tables below show the full results.
Table 2: Five-year GDP forecast accuracy by model, measured by mean absolute percentage error where lower is better, tested on United Arab Emirates data from 1975 to 2023.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Table 3: One-year GDP forecast accuracy by model, measured by mean absolute percentage error where lower is better, tested on United Arab Emirates data from 1975 to 2023.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Figure 1: Average forecast error by model, five-year and one-year horizons, from Tables 2 and 3.
Because a regression model predicts GDP from other factors, so it is only as good as the forecasts of those factors. The paper links GDP to drivers like oil production, exports and interest rates. To test the model fairly, it ran two versions: one fed the real, known values of those drivers, called the perfect prediction, and one fed rough estimates, called the naive prediction. The gap between the two was large, which shows the model’s weakness and its strength at once. Given good inputs it is accurate and consistent for short-term work; given poor inputs it drifts. So it suits a user who has a reliable source for the driving forecasts, and less so one who does not.
Because the two best models are backward-looking, which makes them strong on accuracy but blind to what is coming. ARIMA and VAR learn from the past and project it, so they are accurate when the future resembles the past, but they cannot reflect a change you expect ahead, such as a policy shift or an oil-price move. A linear regression model can, because you can feed it your own view of the drivers, which is exactly why it suits short-term, scenario-based decisions. Running both gives a decision-maker the best of each: ARIMA’s long-range accuracy and regression’s short-range flexibility.
Figure 2: Forecasts against the actual value of GDP for the latest five-year window, in Million AED at 2014 constant prices (the source paper’s final figure).
It might, and the paper is careful to flag this rather than ignore it. GDP fell sharply in 2020, and it is not yet clear whether that was a temporary shock or a permanent change in the pattern of the data, what statisticians call a structural break. If it was a permanent change, all these models would need re-testing and re-estimating once enough post-pandemic years are available, placing more weight on the newer behaviour. The paper treats its results as sound for now while being honest that they should be revisited as the data grows.
The sections below set out how the models were built and tested, for readers who want the full method behind the results above. The complete paper, with every equation and table, is available through the links in the references.
The study drew on a dataset of 426 variables from more than 20 sources, ranging from broad measures like inflation and unemployment to sector-specific ones like oil production, new business licences and hotel guests. Sources included the UAE Federal Competitiveness and Statistics Centre, the Statistics Centre of Abu Dhabi, the Central Bank of the UAE, the International Monetary Fund, the World Bank and Bloomberg. Real GDP data ran annually from 1975 to 2023, in constant 2014 prices and local currency. To test each model fairly, the study used a rolling forecast: it trained the model on the earlier years, forecast the next period, then added the real result and forecast again, repeating across a testing window. This measures how a model would have performed in practice, year after year, rather than on a single lucky forecast.
ARIMA has three settings, written ARIMA(p,d,q), that control how much past data, how much differencing, and how many past errors the model uses. Choosing them well is the whole task. The study first checked whether GDP was stationary, meaning its statistical behaviour is stable over time, using autocorrelation plots and the Augmented Dickey-Fuller test. GDP was not stationary in its raw form and needed differencing. The study then estimated nine candidate ARIMA specifications and compared them using two standard scoring measures, the Akaike and Bayesian Information Criteria, which reward accuracy while penalising needless complexity. The specification ARIMA(1,2,1) scored best on both and was adopted. The three plots below are the diagnostic tools used to reach that choice.
Figure 3: Autocorrelation of United Arab Emirates GDP. The slow, steady decline across years signals that GDP is not stationary in its raw form and needs differencing.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
The Augmented Dickey-Fuller test on the GDP level, reproduced from the source. Interactive on the live page: hover the box for the plain-language reading.
Figure 4: Autocorrelation of GDP after differencing once. The pattern settles closer to zero, indicating the differenced series is closer to stationary.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
The Augmented Dickey-Fuller test on the first difference of GDP, reproduced from the source. Interactive on the live page: hover the box for the plain-language reading.
Figure 5: Partial autocorrelation of GDP, used to judge how many autoregressive and moving-average terms the model needs.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Table 4: The nine ARIMA specifications tested, scored by the Akaike and Bayesian Information Criteria where lower is better; the chosen specification, ARIMA(1,2,1), scored best on both.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
VAR tracks GDP alongside other economic measures and treats them all as moving together, so the first task is choosing which measures to include. Following an approach used for Switzerland and the euro area, the study used the gross value added of each economic sector, meaning how much each sector contributes to output. It then ran a correlation check of every sector against GDP and kept only those correlating above 20 percent, to avoid cluttering the model with weak signals. Mining and quarrying, which includes oil and gas, correlated most strongly at 0.64, followed by wholesale and retail trade at 0.54 and manufacturing at 0.50. The study estimated two versions, VAR(1) and VAR(2). VAR(1) scored better on one measure and VAR(2) on the other, so both were carried into the forecast tests, where VAR(2) proved the more accurate and consistent.
Table 5: How each economic sector correlates with GDP, the screen used to choose the VAR model’s inputs (the source paper’s TABLE VI); sectors below the 20 percent threshold were dropped.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Table 6: The two VAR specifications scored by the Akaike and Bayesian Information Criteria, where lower is better (the source paper’s TABLE VII).
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
The regression model links GDP to a set of economic drivers, so the task is choosing the right drivers and confirming the relationship holds. The study started from nine candidate variables chosen for their known links to GDP, including unemployment, inflation, interest rates, oil production, oil consumption, the global oil price, exports and imports. It checked for multicollinearity, the problem where two inputs move so closely together that the model cannot tell their effects apart, and found only exports and the global oil price were highly correlated, at 84 percent. Through repeated testing it narrowed the model to the variables with the most explanatory power. The final models linked GDP growth to interest rates, oil production, exports and imports, and explained a majority of the variation in GDP growth, but their forecast accuracy depended heavily on the quality of the input forecasts, as the results above showed.
Table 7: The nine candidate variables considered for the regression model, with the reason each was selected (the source paper’s TABLE VIII).
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Figure 6: The estimated regression Model 1
Figure 7: The estimated regression Model 2, dropping the interest-rate lag and adding imports
The paper’s model choice sits on a decade of comparable studies. A systematic review of ten ARIMA papers across eight countries found the family effective wherever it was tried, with the winning specification differing by economy, which is the case for tailoring a model to the United Arab Emirates rather than importing one. And when other authors have raced the same families head to head, the winner has switched with the economy: a VAR won for Albania and an ARIMA for Madagascar. The three tables below carry those results as published.
Table 8: The ten ARIMA studies reviewed by Muma and Karoki, with the specification found best for each economy (the source paper’s TABLE I).
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Table 9: Shahini and Haderi’s comparison of four models for Albania’s quarterly GDP, where the VAR won on every metric (the source paper’s TABLE II).
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Table 10: Andrianady’s comparison for Madagascar’s GDP, where the ARIMA won instead (the source paper’s TABLE III), the counter-example showing the best model changes with the economy.
Source: MCC Economics, full paper on the Munich Personal RePEc Archive.
Master financial modeling with our guide - simplified insights for effective planning, structure, and future adaptability
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Block quote
Ordered list
Unordered list
Bold text
Emphasis
Superscript
Subscript
Empower UK government departments with structured case studies, optimizing public resources effectively.
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Block quote
Ordered list
Unordered list
Bold text
Emphasis
Superscript
Subscript
I was delighted that MCC's work was completed on time, and within budget, helping us deliver important changes and improvements, to the benefit of our stakeholders. MCC's report is published on the CCC website.
I am delighted to recommend MCC Economics. Specifically, I worked closely with PJ, who helped us with our Nuclear and CCUS projects. PJ helped us develop new policies and answer questions from our stakeholders. His support helped us deliver important changes and improvements, to the benefit of our stakeholders.
MCC Economics has helped us better understand the most important issues for our stakeholders, including: charges, shareholder returns, debt payments and inflation impacts.
I am delighted to recommend PJ and his team at MCC Economics. We've been working together on National Policy Statements to help meet net zero targets for 2030 and 2050. We initially appointed MCC Economics to support us on offshore wind consultation analysis and have recently reappointed MCC Economics to undertake a larger consultation analysis role across all sectors, including hydrogen, CCUS and networks. I can confirm that PJ and his team have shown excellent spreadsheet skills, alongside very good project management, planning and analysis skills, helping us deliver important changes, and continuous improvements, to the benefit of our stakeholders.
I am delighted to recommend PJ and his team from MCC Economics. They helped us with our price controls for Heathrow airport and for NATS (En Route) plc (the air traffic services provider). Specifically, the MCC team helped us deliver important changes and improvements to our financial models and supporting policy documents, to the benefit of our stakeholders.
I am delighted to confirm that I worked with PJ on a retail project in 2015. The project helped stakeholders understand electricity costs and charges. Specifically, the project helped us explain to stakeholders, internally and externally, why electricity charges differed across the regions (GB, NI & Ireland). PJ was a key member on the project team, which helped deliver changes and improvements in the understanding of energy retail.
.png)
Evidence-based analysis for transparent, defensible, and effective decision-making.
© All rights reserved – MCC Economics 2026.
.jpg)
