Every economic decision — whether to raise interest rates, approve a budget, or extend credit — is made using data that describes the past. Not yesterday. Often last quarter. By the time an institution knows what the economy was doing, the economy has already moved on. This is not a new problem. But in a world where geopolitical shocks, trade disruptions, and market moves happen in hours, it is becoming an expensive one.
Nowcasting is how economists are closing that gap. And AI is changing what is possible.
The term nowcasting was borrowed from meteorology, where it refers to the practice of predicting current and near-term weather conditions using real-time observations rather than historical models alone. In economics, it refers to estimating the current state of the economy, including GDP growth, inflation, trade volumes, and employment, before official data is available by using high-frequency indicators that are released sooner.
The intuition is straightforward. If you want to know what GDP growth looks like this quarter and official figures won’t arrive for another two months, you look for signals that are available now and that have a reliable relationship with the thing you are trying to measure. Electricity consumption correlates with industrial activity. Port traffic correlates with trade volumes. Credit card spending correlates with consumer demand. Taken together, these signals can give you a surprisingly accurate estimate of where the economy is right now and not where it was last quarter.
For those making critical economic decisions — from central bankers setting interest rates to finance ministers designing fiscal policy and bank risk teams evaluating credit conditions — this real-time insight into the economy is essential. It enables decisions based on today’s economic reality rather than yesterday’s data.
This is the same problem I explored in my previous article on geopolitical risk and AI models — the gap between what a model was trained on and the world it is being asked to describe. Nowcasting is, in some sense, the institutional response to that gap: a systematic attempt to keep economic assessment as close to real time as possible.
Before AI entered the picture, nowcasting relied on two main econometric tools that are worth understanding because they remain the benchmark against which ML approaches are evaluated.
The first is the Dynamic Factor Model (DFM), which, rather than analyzing hundreds of economic indicators independently, assumes that their movements are driven by a small number of common underlying factors, such as the business cycle, global demand, and financial conditions. By estimating these common factors from a large panel of monthly indicators, the model can produce a GDP nowcast that updates as each new data release arrives. This approach is widely used by central banks, including the Federal Reserve Bank of New York for its GDP Nowcast and the European Central Bank for its nowcasting models.
The second is the Bridge Equation — a simpler approach that directly regresses quarterly GDP on monthly indicators. As new monthly data become available during the quarter, the model updates its estimate of quarterly GDP by combining the information from the months observed so far.
Both approaches are grounded in sound econometric theory and have strong track records. They are transparent, interpretable, and produce uncertainty estimates — properties that matter enormously in a policy context where you need to explain your forecast to a minister or a board of governors, not just produce a number.
But these models have limitations. They require careful variable selection, struggle to capture complex nonlinear relationships, and are not well suited to incorporating unconventional, unstructured data such as satellite imagery, shipping signals, social media sentiment, and web-scraped prices — that are now available in enormous quantities and that contains real economic signal.
This is where machine learning enters.
The past three years have seen a significant expansion in the types of data and models being used for economic nowcasting. Three developments in particular stand out.
The first is ** satellite data**. In
The same IMF team published a paper earlier in May 2025 showing that satellite-based vessel tracking data — specifically AIS signals broadcast by cargo ships — can produce a timely indicator of global merchandise trade within seven working days after the end of a month. For context, official trade statistics typically take four to six weeks. The satellite-based model does not replace official statistics. It provides an early warning signal (a nowcast ) that policymakers can act on while they wait for the official figure to arrive.
The second development is the use of ** large language models **for inflation nowcasting. Project Spectrum — a collaboration between the Bank for International Settlements, the Deutsche Bundesbank, and the European Central Bank published in February 2026 — demonstrated that combining text embeddings with machine learning algorithms can efficiently classify product descriptions from scanner and web-scraped data into official price indices. This allows central banks to track price developments at a granularity and timeliness that traditional CPI surveys cannot match. The ECB now uses LLMs to nowcast and forecast inflation, combining text data with ML techniques to quantify risks in the global economy in near-real time.
The third development and the most nuanced is the evidence on** when ML outperforms traditional econometric models, and when it does not**.
A December 2025 IMF working paper conducted the most comprehensive evaluation to date by comparing dozens of models across multiple countries. The research concluded that traditional econometric models — particularly Bridge Models and Dynamic Factor Models — tend to outperform complex ML algorithms for GDP nowcasting. Among ML approaches, linear models like Lasso and Elastic Net perform best, even outperforming traditional models when GDP data is long and high-frequency indicators are rich.
The reason is straightforward when you think about it. Complex non-linear ML algorithms like gradient boosting and neural networks are prone to overfitting on the relatively short GDP time series available in most countries. A model that fits the training data very well but generalizes poorly is exactly what you do not want in a nowcasting context where out-of-sample performance is the only thing that matters.
This does not mean AI adds nothing. It means the value of AI in nowcasting lies in the new data it can handle — satellite imagery, vessel tracking signals, web-scraped prices — not necessarily in replacing the econometric models that process that data. The economist who knows when to use a regression and when to use a random forest is more valuable than one who defaults to the most complex tool available.
The example below simulates a simple nowcasting setup — using a high-frequency monthly indicator to estimate a quarterly GDP figure before the official release arrives. Think of it as a stripped-down bridge equation implemented in Python.
import numpy as npimport pandas as pdfrom sklearn.linear_model import LinearRegressionfrom sklearn.metrics import mean_absolute_errorimport matplotlib.pyplot as pltnp.random.seed(42)# Simulating quarterly GDP growth (%) — available with a 6-8 week lag# Represents 5 years of quarterly data (20 quarters)# Notice how COVID hit in 2020, recovery in 2021, rate tightening in 2022quarters = pd.date_range(start='2019-01-01', periods=20, freq='QE')gdp_growth = np.array([2.1, 1.8, 2.3, 2.0, # 2019 1.9, -8.5, -7.2, 3.4, # 2020 — COVID shock 4.2, 3.8, 3.1, 2.9, # 2021 — recovery 2.4, 2.1, 1.8, 1.5, # 2022 — rate tightening 1.8, 2.0, 2.2, 1.9]) # 2023# Simulating a high-frequency monthly indicator — e.g. port throughput index# Available in near-real time and is correlated with GDP# Watch how it collapses in early 2020 then recovers — just like GDP didmonths = pd.date_range(start='2019-01-01', periods=60, freq='ME')port_index = np.array([98, 102, 100, # Q1 2019 97, 99, 98, # Q2 2019 101, 103, 102, # Q3 2019 99, 101, 100, # Q4 2019 98, 96, 60, # Q1 2020 — COVID collapse 45, 48, 50, # Q2 2020 65, 75, 85, # Q3 2020 — recovery begins 90, 95, 98, # Q4 2020 102, 105, 108, # Q1 2021 107, 109, 110, # Q2 2021 108, 106, 105, # Q3 2021 104, 103, 102, # Q4 2021 100, 99, 98, # Q1 2022 97, 96, 95, # Q2 2022 94, 93, 92, # Q3 2022 91, 90, 91, # Q4 2022 92, 94, 95, # Q1 2023 96, 97, 98, # Q2 2023 99, 100, 101, # Q3 2023 100, 99, 98]) # Q4 2023# Average monthly indicator to quarterly frequencyquarterly_port = port_index.reshape(20, 3).mean(axis=1)# Building the nowcasting model: training on first 16 quarters (2019-2022)# Nowcasting the last 4 quarters (2023) — pretending we don't have official GDP yettrain_X = quarterly_port[:16].reshape(-1, 1)train_y = gdp_growth[:16]test_X = quarterly_port[16:].reshape(-1, 1)test_y = gdp_growth[16:]model = LinearRegression()model.fit(train_X, train_y)nowcast = model.predict(test_X)mae = mean_absolute_error(test_y, nowcast)print("Nowcast vs Actual GDP Growth (2023):")for i, (actual, predicted) in enumerate(zip(test_y, nowcast)): print(f" Q{i+1} 2023 — Actual: {actual:.1f}% Nowcast: {predicted:.1f}%")print(f"\nMean Absolute Error: {mae:.2f} percentage points")# Visualisefig, ax = plt.subplots(figsize=(12, 5))ax.plot(range(20), gdp_growth, 'b-o', label='Actual GDP Growth', linewidth=2)ax.plot(range(16, 20), nowcast, 'r--s', label='Nowcast (before official release)', linewidth=2, markersize=8)ax.axvline(x=15.5, color='gray', linestyle=':', linewidth=1.5, label='Nowcast begins')ax.set_xticks(range(20))ax.set_xticklabels(['Q1\n19','Q2','Q3','Q4','Q1\n20','Q2','Q3','Q4', 'Q1\n21','Q2','Q3','Q4','Q1\n22','Q2','Q3','Q4', 'Q1\n23','Q2','Q3','Q4'], fontsize=8)ax.set_ylabel('GDP Growth (%)')ax.set_title('Nowcasting GDP: High-Frequency Port Data vs Official Release')ax.legend()ax.grid(True, alpha=0.3)plt.tight_layout()plt.savefig('nowcast_gdp.png', dpi=150)plt.show()
Results below :
Nowcast vs Actual GDP Growth (2023):
Q1 2023 — Actual: 1.8% Nowcast: 1.2%
Q2 2023 — Actual: 2.0% Nowcast: 1.9%
Q3 2023 — Actual: 2.2% Nowcast: 2.6%
Q4 2023 — Actual: 1.9% Nowcast: 2.3%
Mean Absolute Error: 0.38 percentage points
The blue line is official GDP growth — the number that gets published six to eight weeks after each quarter ends. The red dashed line is our nowcast — what we estimated using only port data, before the official figure was available. The grey dotted vertical line marks the point where we stopped training and started nowcasting.
If the red and blue lines are close in 2023, our port-based signal is doing its job — giving us a reasonable read on the economy before the official statistics confirm it. That is the entire point of nowcasting. Not perfect precision, but useful directional intelligence, available weeks earlier than the official release.
The UAE presents a particularly interesting nowcasting context for three reasons.
First, as a small, open and trade-dependent economy, the UAE feels external shocks faster than most. An oil price move, a shipping disruption, a drop in tourism bookings — these transmit into domestic conditions within weeks, not quarters. By the time an official GDP figure is published, it may already be describing a world that has moved on.
Second, Jebel Ali Port — one of the largest container ports in the world and the busiest in the Middle East — produces AIS vessel tracking data that is ideal for nowcasting UAE trade and economic activity. Port throughput at Jebel Ali is not just a trade indicator — it is a real-time proxy for UAE non-oil GDP activity, given the port’s centrality to the UAE’s re-export economy.
Third, the UAE’s non-oil economic diversification agenda — central to Vision 2031 — requires more granular, more timely measurement of where non-oil growth is actually coming from. Is tourism growing? Are financial services expanding? Official statistics answer these questions annually. Credit card transaction data, electricity consumption figures, hospitality booking platforms, and real estate registries could answer them in near-real time.
The tools exist. The data exists. The question for UAE institutions is not whether to build these capabilities — it is how quickly.
Intellectual honesty requires acknowledging what nowcasting cannot do.
It cannot predict shocks. COVID did not show up in any leading indicator before March 2020. The Red Sea disruption started with a geopolitical event that no economic time series could have anticipated. Nowcasting tracks the economy along its current trajectory — they are not designed to forecast structural breaks of the kind I explored in my previous article.
It cannot replace judgment. A nowcast model that estimates UAE GDP growth at 4.2% for the current quarter is producing a statistical estimate, not a forecast. It does not know about the policy changes that were announced last week, the trade agreement that is being renegotiated, or the investment project that was cancelled. These are things a human economist with institutional knowledge needs to incorporate.
And more sophisticated AI does not always mean better results. The right tool depends on the data availability, series length, and the specific nowcasting horizon — not on which model sounds most impressive. For many GCC economies where historical data is shorter, a well-specified econometric model often outperforms a complex machine learning algorithm.
What AI genuinely adds is breadth — the ability to incorporate types of data that traditional models cannot handle: unstructured text, satellite imagery, vessel tracking signals, web-scraped prices. Not a replacement for econometric rigor, but an expansion of the information set that rigor can be applied to.
Official statistics tell you where the economy was. Nowcasting tells you where it is. That distinction — small in theory, significant in practice — is becoming one of the most valuable capabilities a policy institution can have.
AI has expanded what is possible. Satellites, vessel signals, and transaction data are giving economists access to signals that did not exist a decade ago. But the questions that matter — which signals are reliable, which relationships are stable, how to account for structural breaks, how to communicate uncertainty to decision-makers — remain fundamentally economic questions.
The gap between data and decision is where economic analysis has always lived. Nowcasting is simply the newest instrument for closing it.
Eram Tafsir is an Applied Economist and Quantitative Analyst with experience in econometrics, ML modeling, and macroeconomic analysis. She writes about the intersection of economics, data science, and AI.
Nowcasting in the Age of AI: How Real-Time Data Is Changing the Way We Read the Economy was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.