AI and machine-learning models generally outperform traditional time-series methods in forecasting economic indicators, though data gaps and study heterogeneity limit firm conclusions; early tests show ChatGPT and Google Gemini generate competitively accurate predictions for EU GDP and unemployment.

Main Finding

AI and machine‑learning methods (especially neural networks and sequence models, often in hybrid form) improve accuracy and timeliness in forecasting and nowcasting macroeconomic and financial indicators compared with many traditional time‑series models. However, data availability/quality and bias/privacy concerns remain critical constraints. Empirical evidence on using large language model (LLM) chatbots for direct forecasting is still limited; the authors’ own small empirical tests with ChatGPT 5.2 and Gemini AI Pro on EU GDP (current prices) and unemployment rate produced “good” MAPE‑based performance, but chatbot work in the literature is mostly conceptual.

Key Points

• Scope: Systematic literature review using Web of Science (WoS) that produced a final sample of 111 papers (from an initial 399) on forecasting/nowcasting with AI and chatbots in economics.
• Time trend: Strong increase in publications after 2019, peaking in 2024; top outlet: Journal of Forecasting.
• Study mix: 94 empirical papers, 17 conceptual/review papers.
• Common target variables: stock market indices, electricity & oil prices, cryptocurrency, exchange rates, bankruptcy/default, GDP, unemployment rate.
• Most used AI methods: Artificial Neural Networks (ANN), Long Short‑Term Memory (LSTM), Gated Recurrent Unit (GRU), Support Vector Machine (SVM). Hybrid models frequently use optimizers/feature methods like Particle Swarm Optimization (PSO), Genetic Algorithms (GA), Random Forest (RF).
• Main empirical conclusion across studies: AI/ML models (and hybrids) tend to outperform traditional time‑series methods (e.g., ARIMA) in many forecasting tasks, conditional on data availability and proper tuning.
• Chatbots / LLMs: Literature is mostly conceptual; few empirical comparisons. The authors ran illustrative forecasts with ChatGPT 5.2 and Google Gemini Pro and report favorable MAPE results for EU GDP and unemployment forecasts.
• Risks/limitations: data availability and frequency limitations; bias (e.g., anchoring effects found in some LLMs), interpretability, data privacy, and potential ethical/security issues.
• Noted influential works: Petropoulos et al. (2022) for forecasting theory and methods; Goodell et al. (2021) and Barboza et al. (2023) for finance/AI reviews.

Data & Methods

• Search strategy: Web of Science searches combining keywords (“forecasting”, “nowcasting”, “artificial intelligence”, “chatbot”, “chat”, “bot”) with WoS category restricted to “Economics” for some queries; all‑fields queries used otherwise.
• Screening and selection:
  • Initial hits: 399 papers.
  • Exclusions: conference papers (85), inaccessible full texts (7), duplicates (6).
  • Abstract screening removed papers not focused on macroeconomic indicators (176) and papers using non‑public toy/data‑package datasets (14).
  • Final sample: 111 papers (94 empirical, 17 conceptual/review).
• Coding/analysis:
  • Papers classified by publication year, journal, citation counts, whether conceptual/empirical, dependent variable, and AI method used.
  • Empirical studies summarized by domain (stock indices, energy prices, cryptocurrencies, FX, bankruptcy, GDP, unemployment).
• Authors’ empirical chatbot test:
  • Selected ChatGPT 5.2 and Gemini AI Pro.
  • Forecast targets: EU GDP (current prices) and EU unemployment rate.
  • Evaluation metric: mean absolute percentage error (MAPE); chatbots “showed good predicting performances” per MAPE (exact numeric MAPE values not reported in the provided excerpt).
• Limitations of the review methodology noted by authors:
  • Use of a single database (WoS) and restrictions on category for some queries.
  • Exclusion of conference papers and studies using private/non‑public datasets.
  • The chatbot evidence base is sparse and mainly conceptual.

Implications for AI Economics

• Forecast quality and timeliness: AI/ML methods can materially improve short‑term forecasting and nowcasting for many economic/financial series when sufficient, high‑quality data are available—this can aid policymakers, central banks, firms, and investors for faster, better informed decisions.
• Data infrastructure & governance: Gains from AI hinge on access to timely, granular, and privacy‑respecting data. Governments and statistical agencies should prioritize safe data sharing, standardization, and investments in real‑time data pipelines to unlock forecasting value.
• Policy and regulation: AI strategies at the governmental level should address (a) data privacy and ownership, (b) standards for model validation and transparency, and (c) measures to reduce model bias and manipulation risk (e.g., anchoring or other LLM biases).
• Chatbots/LLMs as forecasting tools: LLMs are promising (easy access, flexible interfaces) but under‑evaluated empirically for numeric economic forecasting. Before adoption in policy settings, systematic benchmarking against established econometric and ML models, robustness checks, and explanation/uncertainty outputs are needed.
• Capacity and literacy: Users (analysts, policymakers) must develop skills to interpret AI‑generated forecasts, understand limits and biases, and combine model outputs with domain knowledge and judgment.
• Research agenda: More empirical work comparing LLMs, classical ML, and econometric approaches on standard datasets; standardized evaluation metrics (including probabilistic forecasts and economic loss functions); examination of hybrid pipelines (LLM for feature extraction/text signals combined with time‑series ML); and reproducible benchmarks with public data.

Limitations to bear in mind: results reflect the WoS‑based sample and authors’ own limited chatbot experiments; exact numeric results for the chatbot tests are not reported in the excerpt.