First use of AI in inflation forecasting at the CNB

1. Introduction

In this blog article, we ask whether and how the use of artificial intelligence (AI) could enhance inflation forecasting in the Czech Republic. As this is the first time the Czech National Bank (CNB) has explored this approach, we are not yet seeking to draw definitive conclusions. This is a test and demonstration of the initial possibilities for using AI in inflation forecasting at the CNB.

The advantage of AI models is that they operate independently of any specific economic theory. The paradigms of New Keynesian economics and monetarism play no role here.  These models represent the distilled experience of millions of people and learn from petabytes of data. They offer a fresh perspective on inflation forecasting – one that may be less constrained by theory and more data-driven and practically oriented. AI allows data analyses on an unprecedented scale and depth, from historical statistics to real-time trends. However, the downside is that AI models operate as a black box and may lie or exhibit errors such as “hallucinations” or inaccurate interpretations during backtesting. Despite this, we believe that AI offers a promising way to complement and extend existing approaches to inflation forecasting at the CNB.

For inflation forecasting, we will use two main AI subcategories: large language models (LLMs), specifically OpenAI o1 and Grok 2 for medium-term inflation forecasting, and machine learning models (a detailed overview of these models can be found in the Appendix). We will use the embeddings method, a machine learning approach, to improve inflation nowcasting (i.e. estimating current, as yet unpublished inflation).

2. First use of AI for medium-term inflation forecasting at the CNB

The first approach involves using large language models (LLMs) for medium-term inflation forecasting. Methodologically, we build on the work of Faria-e-Castro and Leibovici (2024). Newer LLMs have access to vast amounts of up-to-date data, including macroeconomic statistics, reports and economic outlooks. These can be used to generate independent inflation forecasts. You just ask the model what inflation it expects for the following year, and it gives you an answer.

Our focus is on assessing the accuracy of LLMs in forecasting inflation. For this experiment, we used two state-of-the-art models: OpenAI o1 from OpenAI and Grok 2 from x.AI. The prompt for both models was to forecast year-on-year inflation in the Czech Republic one year ahead. For each quarter from 2019 Q1 to 2024 Q3, we told the model to do the following:

“Assume you are at time 𝝉 with no access to any later information. Please provide your best independent forecast of year-on-year seasonally unadjusted CPI inflation in the Czech Republic for the next year. Include numerical values for these forecasts. Base your projections only on the information available as of 𝝉.”

Here, 𝝉 represents the individual quarters from 2019 Q1 to 2024 Q3.

Using this query, we obtained AI-generated inflation forecasts for the following year. Now, let’s examine how these AI models performed compared to actual observed inflation, financial market analysts’ forecasts, and forecasts using the CNB’s core model. As the benchmark, we use the root mean square error (RMSE), which measures the mean prediction error. A higher RMSE indicates a greater deviation between the forecasted and actual values.

Table 1 summarises the overall forecasting errors of the AI models and analysts. Over the entire period, OpenAI o1 demonstrated the highest accuracy, with an RMSE of 5.28, outperforming the CNB’s core model (RMSE = 5.48). OpenAI o1 performed better in forecasting inflation while it was rising, while the CNB’s core model was more accurate as inflation returned to the 2% target. In periods of stable inflation, Grok 2 was the most accurate model.

Table 1 – One-year-ahead inflation forecasting errors based on RMSE
The table presents the RMSE values of the one-year-ahead inflation forecasts from the two AI models tested (OpenAI o1 and Grok 2), financial market analysts, and the CNB’s core model, compared to actual observed CPI inflation. The forecasts were produced between 2019 Q1 and 2023 Q3, corresponding to the inflation outcomes from 2020 Q1 to 2024 Q3. CPI inflation measured quarterly was used for the comparison.

Period
(start–end)
Overall
(2020
Q1–2024 Q3)
Stable inflation
(2020
Q1–2021 Q1)
High inflation
(2021
Q2–2023 Q1)
Return to 2% target
(2023
Q2–2024 Q3)
OpenAI o1 5.28 1.09 7.94 1.81
Grok 2 6.19 0.50 9.05 3.45
Financial market analysts 6.63 0.96 9.99 2.31
CNB core model 5.48 0.79 8.41 0.50

Source: authors, CNB

Charts 1 and 2 provide a detailed comparison of the results of the AI models and financial market analysts. The charts illustrate the one-year-ahead inflation forecasts from the AI models over the entire period under review (2019 Q1–2024 Q3) and compare them with actual observed CPI inflation. The charts highlight three key elements: (i) red dots representing the inflation forecasts from the AI models, (ii) a blue line showing actual observed year-on-year inflation, and (iii) yellow lines depicting the forecasts from financial market analysts. Additionally, grey lines in the chart show the inflation path predicted by OpenAI o1 for different periods.

To ensure a correct comparison, we shifted the financial market analysts’ forecasts forward by one year. This ensures that both forecasts in Charts 1 and 2 represent inflation projections made 12 months in advance. If a forecast (the red dots or the yellow line) aligns with actual observed CPI inflation (the blue line), it indicates an accurate prediction of the inflation path. If the forecast falls below the blue line, inflation was underestimated; if it is above, inflation was overestimated.

Chart 1 – One-year-ahead inflation forecast with OpenAI o1 and observed inflation
The chart shows the forecasts of observed inflation (12-month-ahead outlook) generated by OpenAI o1 (red dots) and financial market analysts (yellow line). These forecasts were made 12 months in advance of the given period. For comparison of accuracy, the chart includes actual observed CPI inflation (blue line) at a monthly frequency. Additionally, the grey line represents the inflation path predicted by OpenAI o1, illustrating its estimated path over time.

 Chart 1 – One-year-ahead inflation forecast with OpenAI o1 and observed inflation

Source: authors, CNB

Chart 1 illustrates the inflation forecast generated by the AI model OpenAI o1. During the period of relatively stable inflation (2019–2020), the forecasts from OpenAI o1 and financial market analysts were largely aligned, both anticipating a swift return of inflation to the CNB’s target. However, the situation changed dramatically in 2021–2022, when inflation surged – an increase that neither method managed to predict accurately.

However, OpenAI o1 started signalling the possibility of further inflation growth earlier, with its forecasts indicating a persistent increase for most of 2021. In contrast, financial market analysts’ forecasts suggest that they considered the higher inflation to be a temporary deviation, expecting a swift return to the target. However, neither the analysts nor OpenAI o1 were able to accurately estimate the actual level of inflation during this period. Overall, the results suggest that OpenAI o1 could serve as an indicator of future inflation growth.

Let’s now focus on the second AI model tested – Grok 2. Chart 2 shows that Grok 2 more closely follows the financial market analysts’ forecasts during periods of both stable inflation and rising inflation. However, its forecasts are not entirely identical, which may suggest that the model did not “secretly” draw on the analysts’ predictions. This is particularly evident at the end of 2023 and the start of 2024, when Grok 2 expected inflation to remain higher for longer, while financial market analysts – in line with actual developments – anticipated a rapid decline towards the 2% target.

Chart 2 – One-year-ahead inflation forecast using Grok 2 and observed inflation
The chart shows the forecasts of observed inflation (12-month outlook) generated by Grok 2 (red dots) and financial market analysts (yellow line). These forecasts were made 12 months in advance of the given period. For comparison of accuracy, the chart includes actual observed CPI inflation (blue line) at a monthly frequency. Additionally, the grey line represents the inflation path predicted by Grok 2, illustrating its estimated path over time.

Chart 2 – One-year-ahead inflation forecast using Grok 2 and observed inflation

Source: authors, CNB

The differences between Grok 2 and OpenAI o1 suggest that AI’s ability to forecast inflation can vary significantly depending on the specific model architecture. OpenAI o1 is better at assessing the available information because it processes it repeatedly before answering a question. By contrast, Grok 2 responds instantly, without further consideration. A fundamental issue with both models, however, is that they function as black boxes – we do not know exactly how they arrive at their conclusions. For example, we do not know what assumptions they make about interest rates, exchange rates and so on. Additionally, different generations of models may reason in different ways.

International comparison and other medium-term inflation forecasting methods

Regarding international comparisons of medium-term inflation forecasting methods, Fed analysts Faria-e-Castro and Leibovici (2024) also showed that large language models can estimate inflation trends more accurately than professional analysts. For their analysis of the 2019–2023 period, they used Google’s PaLM model, which was more accurate in most years and across nearly all time horizons. However, a recurring issue with these forecasts is that it is not entirely clear how AI arrives at its predictions, as it operates as a black box. Nevertheless, AI-based forecasts could serve as a valuable supplement to existing inflation expectations surveys, such as those conducted among financial market participants, households and businesses.

Machine learning models offer another approach to medium-term inflation forecasting. Advanced methods include decision trees, such as random forests, which can process large numbers of input variables, including commodity prices, interest rates and consumer confidence. Due to their high accuracy, gradient boosting machines – including XGBoost and LightGBM – are particularly well suited for forecasting short-term inflation pressures. This remains an area for further research. The ECB already uses these models as a supplementary tool alongside its core forecasting model. The main advantage of these methods is their ability to capture non-linear relationships and inflation dynamics, which is especially important during sudden shocks, such as Covid-19 and the subsequent rise in prices. Lenza et al. (2023) at the ECB use the latest machine learning method, quantile random forests, which not only forecast future inflation but also quantify the uncertainty of estimates. This enables central bank boards to better assess inflation risks and obtain a detailed overview of potential inflation scenarios, including extreme ones. Machine learning models are also being explored by institutions such as the BIS (Kohlscheen, 2021), the IMF (Liu et al., 2024), the Bank of England (Joseph et al., 2022) and the Central Bank of Brazil (Boaretto & Medeiros, 2023).

3. First use of AI in inflation nowcasting at the CNB

Another potential application of artificial intelligence is nowcasting, which involves estimating current, unpublished inflation data. At the CNB, we already monitor daily food price developments using millions of automatically collected data points from the internet (web scraping).

However, central banks worldwide do not fully utilise the potential of the available data. One of the main limitations is that some products are known only by their marketing name, making it difficult for traditional methods to accurately classify them into the correct category of the consumer basket.

Currently, products are categorised using an ex-ante approach, where prices are assigned to consumer basket categories based on the e-shop section in which they are listed. However, this method has its limitations – for example, baby food may be sold in the “For children” category, making it difficult to classify it within the food section of the consumer basket. An alternative approach involves creating a custom product library based on keywords, but this method is highly time-consuming and lacks flexibility.

The embeddings method, which forms the backbone of every large language model, may provide a solution to this challenge. The embeddings method works by converting product names into numerical representations (vectors) that a computer can process. Products with similar meanings will have more similar numerical vectors, placing them closer together in the embedding space. This is because embeddings not only connect words based on linguistic similarity, but also capture their meaning. For example, a restaurant and a café are both service establishments, while a restaurant and a tractor have little in common and will have less similar vectors. This allows us to group texts with similar meanings together.

With embeddings, we can transform the chaos of hundreds of millions of online products into clearly defined consumer basket categories, significantly improving the accuracy and efficiency of inflation forecasts.

Chart 3 illustrates this concept. Consider three products: “Dr. Halíř butter”, “iPhone 16” and “olive oil”. Using AI embeddings, we convert them into numerical vectors. The left side of the chart displays these products in the vector space (each vector number represents an axis, i.e. a dimension). It is already evident that Dr. Halíř butter and olive oil are more similar and therefore closer together, whereas the iPhone 16 is further away as it is less related. The right side of the chart adds two consumer basket categories – “oils and fats” and “mobile phones”. Each of these categories also has a position in the vector space and a certain distance from the products. The shortest distance between each product and category (marked by arrows) indicates the most appropriate category for the product. In this case, butter and olive oil fall into the “oils and fats” category, while the iPhone 16 is classified under “mobile phones”.

Chart 3 – Use of embeddings for product categorisation in consumer baskets
The left chart displays the positions of three products in a three-dimensional vector space generated using embeddings (Dr. Halíř butter, olive oil and iPhone 16). The right chart also shows the positions of two product categories (mobile phones, oils and fats). Arrows indicate the shortest distance between each product and its corresponding category. Products are shown in orange and categories in blue. The axes represent the embeddings vector space, which in this case consists of three dimensions. The axes indicate the proximity of one text to another.

Chart 3 – Use of embeddings for product categorisation in consumer baskets

Source: authors

This process can be repeated for all products whose prices are obtained from the internet. Products such as Dr. Halíř butter and the iPhone 16 can be converted into numerical vectors using advanced AI models. These vectors are then compared with those representing different consumer basket categories, such as “oils and fats” and “mobile phones”. Based on these comparisons, products can be precisely categorised and their prices can be used for nowcasting. The Bank for International Settlements (BIS) has recently started experimenting with this approach in its Spectrum project.

To illustrate this with a practical example using Czech data, we collected thousands of products from the Czech Tesco online store to classify them into specific consumer basket categories using AI and integrate the online data into inflation nowcasting. We analysed each product using the best embeddings available (the text-embedding-3-large model), labelled them using ChatGPT 4o-mini, converted them into numerical representations, and assigned them to consumer basket categories.

Chart 4 – Various products offered by Tesco segmented into consumer basket categories
The chart displays products offered by Tesco in a three-dimensional embeddings vector space (three axes). Initially, each product is represented by 3,072 numbers, meaning it has 3,072 dimensions. To visualise the data, we reduced the number of dimensions to three using principal component analysis (PCA). The three axes in the chart represent the first, second and third principal components of the embeddings dimensions. The axes characterise the similarity of one product to another: the closer the points, the more similar the products. Products within the same consumer basket category are assigned a specific colour: yellow represents “oils and fats” (e.g. butter), green represents “non-alcoholic beverages” (e.g. mineral and flavoured water), orange represents “beer”, and grey represents other categories.

Chart 4 – Various products offered by Tesco segmented into consumer basket categories

Source: products available in Tesco on 17 January 2025

Chart 4 illustrates the positions of the various products in the vector space, where each point represents a single product. Different colours indicate the consumer basket categories to which the products have been assigned. We selected three categories: oils and fats, non-alcoholic beverages and beer. The chart clearly demonstrates that products within the same category are grouped closely together, and even related consumer basket categories are positioned closer to each other.

We also compared the product categorisation using embeddings with the ex-ante method (assigning categories based on the Tesco listing). At a standard level of detail, the two methods matched in 89.2% of cases, demonstrating the strong potential of the embeddings method. Additionally, embeddings proved to be more robust than the ex-ante method, while showing room for further improvement. Table 2 presents a comparison of the embeddings method and the ex-ante method across various levels of consumer basket detail.

Table 2 – Agreement between embeddings and ex-ante methods for product categorisation
The table illustrates the degree of alignment between the two categorisation methods: the AI-driven embeddings method and the ex-ante method, which assigns products based on Tesco’s categorisation. Product categorisation can vary in detail – at the highest level, products are assigned to broad categories such as “food and non-alcoholic beverages”, whereas the lowest level involves highly specific classifications, such as “butter and margarines”. The table shows how well these two methods align in assigning all the products offered by Tesco at each level of consumer basket detail.

Consumer basket detail Example of consumer basket category Category code Agreement between embeddings and ex-ante method
Level 1 (broadest) Food and non-alcoholic beverages E01 99.7%
Level 2 Food E01.1 98.5%
Level 3 Oils and fats E01.15 89.2%
Level 4
(most detailed)
Butter and margarines E01.151 80.1%

Source: authors, CNB

Thanks to embeddings, we can categorise a significantly higher number of items into precisely defined consumer basket categories, allowing us to generate more accurate consumer inflation forecasts.

In terms of international comparison, the BIS, along with the European Central Bank (ECB) and the Bundesbank, launched the Spectrum project (external link) in November 2024 to enhance AI-driven nowcasting. In Poland (Macias et al., 2023), online store data collection has been ongoing since 2009, but only 25% of the data collected is currently used for nowcasting. In Austria (Beer et al., 2024), only 13% of the data collected is utilised, yet improvements in inflation nowcasting are already evident. Meanwhile, the ECB is developing its own research network, Prisma (external link), with which CNB experts have long maintained contact, allowing us to build on the latest findings and apply them in the Czech Republic. AI has the potential to significantly improve how we utilise online data for more accurate nowcasting. Other examples of nowcasting include Knotek and Zaman (2024) and Schnorrenberger et al. (2024).

4. Conclusion, opportunities, and limitations of AI

In this article, we have presented two examples of the application of AI in inflation forecasting – medium-term inflation forecasts using an LLM model and nowcasting using the embeddings method. Our key findings are as follows:

(1) Language models such as ChatGPT and Grok have, in certain periods, been able to generate inflation forecasts that outperformed professional analysts and even the CNB’s own model

Our experiment demonstrated that advanced language models were particularly effective at capturing trend shifts, such as sharp increases in inflation. These models can serve as a valuable tool for identifying potential risks associated with inflation pressures. However, it is important to emphasise that AI cannot yet be considered a full-fledged replacement for traditional forecasting methods. The main challenge remains its black-box nature, making it difficult to determine which factors influenced the predictions. Furthermore, these models exhibit limited reliability when applied to data outside their training set. The forecasts also contain an element of randomness, making them difficult to replicate and increasing the need to have the results checked by experts. Similarly, Fed analysts Faria-e-Castro and Leibovici (2024) found that Google’s PaLM model was more accurate in most years and across almost all time horizons than professional analysts, yet they also identified similar risks in using the current generation of AI models.

(2) AI has the potential to significantly improve inflation nowcasting by accurately classifying millions of online prices into the consumer basket

Our findings indicate that methods such as embeddings can process and categorise vast amounts of data from online stores, leading to more accurate estimates of current inflation pressures. Embeddings is a technique that converts textual information, such as product names found online, into numerical representations. This allows products with the same or similar meaning (e.g. “butter” and “oil”) to be grouped into the appropriate consumer basket category, while dissimilar products (e.g. “butter” and “mobile phone”) remain separated. This method facilitates more efficient sorting of unstructured data and their use in estimating consumer inflation.

(3) The current generation of AI is suited as a support tool for analysts

Despite its limitations, AI provides significant support for analytical work. It accelerates data processing, enables the creation of alternative forecasts and offers additional perspectives on medium-term inflation developments. By doing so, it enhances the quality of analytical processes and supports decision-making based on a broader range of information.

(4) The challenge for AI is to provide medium-term inflation forecasts based on hundreds of macroeconomic indicators – machine learning

At the CNB, the primary forecasting scenario is based on the g3+ DSGE model. However, from 2021 to 2023, this model exhibited larger forecast errors for one-year-ahead inflation predictions (as measured by RMSE) compared to the models used by the Bank of England, the ECB, Sweden’s Riksbank and Norway’s Norges Bank. Since January this year, we have been developing a new semi-structural model as an alternative to the existing primary model. This diversification of the modelling framework follows recommendations made in last year’s external review. In addition to these models, we can incorporate forecasts generated through machine learning, which is better suited to detecting macroeconomic turning points (see the literature review in the Appendix). Machine learning models can identify hidden patterns in historical data and enhance the accuracy of predictions of macroeconomic changes. These methods identify patterns within historical relationships across hundreds of time series. The ECB uses them as a complement to traditional macroeconomic models to verify the robustness of its primary model’s forecasts (Lenza et al., 2023).

(5) A high-quality and extensive data foundation is a necessary condition for any AI-based model

The successful implementation of AI methods requires reliable and high-quality data. Models trained on outputs from previous generations may gradually lose their ability to respond accurately, underscoring the importance of data availability and accuracy (Shumailov et al., 2024). At the CNB, we therefore emphasise data accuracy and accessibility within our public ARAD presentation system and are improving information-sharing processes across the institution through an internal data warehouse. We have also centralised the CNB’s economic research, merging the research activities of the Monetary Department and the Financial Stability Department into a single Research and Statistics Department. Closer integration with statistics enables more efficient use of underlying data and, through closer collaboration between users and data producers, an overall improvement of statistical datasets.

References

Beck, G. W., Carstensen, K., Menz, J.-O., Schnorrenberger, R., & Wieland, E. (2024). Nowcasting consumer price inflation using high-frequency scanner data: evidence from Germany (Working Paper Series No. 2930). European Central Bank.

Beer, C., Ferstl, R., & Graf, B. (2024). Improving Disaggregated Short-term Food Inflation Forecasts With Webscraped Data. https://ssrn.com/abstract=5016579

Boaretto, G., & Medeiros, M. C. (2023). Forecasting inflation using disaggregates and machine learning. https://arxiv.org/abs/2308.11173

Faria-e-Castro, M., & Leibovici, F. (2024). Artificial Intelligence and Inflation Forecasts. Working Paper 2023, 15, Federal Reserve Bank of St. Louis.

Joseph, A., Potjagailo, G., Kalamara, E., Chakraborty, C., & Kapetanios, G. (2022). Forecasting UK inflation bottom up (Staff Working Paper No. 915). Bank of England.

Knotek II, E. S., & Zaman, S. (2024). Nowcasting Inflation. Federal Reserve Bank of Cleveland, Working Paper No. 24-06.

Kohlscheen, E. (2021). What does machine learning say about the drivers of inflation? (Working Paper No. 980). Bank for International Settlements.

Lenza, M., Moutachaker, I., & Paredes, J. (2023). Density forecasts of inflation: a quantile regression forest approach (Working Paper Series No. 2830). European Central Bank.

Liu, Y., Pan, R., & Xu, R. (2024). Mending the Crystal Ball: Enhanced Inflation Forecasts with Machine Learning (Working Paper WP/24/206). International Monetary Fund.

Macias, P., Stelmasiak, D., & Szafranek, K. (2023). Nowcasting food inflation with a massive amount of online prices. International Journal of Forecasting, 39(2), 809–826.

Shumailov, I., Shumaylov, Z., Zhao, Y. et al. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755–759.

Schnorrenberger, R., Venes Schmidt, A., & Moura, G. V. (2024). Harnessing Machine Learning for Real-Time Inflation Nowcasting. De Nederlandsche Bank Working Paper No. 806.


Appendix

A1. Literature review

Large language models and machine learning are used in two key types of inflation forecasting: nowcasting and medium-term inflation forecasting. The latest studies are summarised in Table A.1, which outlines the methods used by various authors and their conclusions.

Table A.1 – Summary of the literature on nowcasting and medium-term forecasting

Field Institution Authors/Project Method Result
Nowcasting BIS Project Spectrum Structuring billions of products using LLMs for nowcasting Project started in November 2024, no deliverables yet
Cleveland Fed Knotek & Zaman (2024) Mixed-frequency models Parsimonious mixed-frequency model provides more accurate nowcasts for headline inflation and competitive nowcasts for core inflation
DNB Schnorrenberger et al. (2024) Machine learning with high-frequency macrofinancial indicators Machine learning improves weekly inflation nowcasts, particularly during inflation after Covid-19
ECB Beck et al. (2023) (PRISMA Project) Use of MIDAS method and machine learning on scanner data Improved headline inflation nowcasting
NBP Macias et al. (2023) Use of SARMAX model on web-scraped prices Improved food price inflation nowcasting
OeNB Beer et al. (2024) Time series models and machine learning for nowcasting web-scraped prices Faster and more detailed availability of food price inflation before official data
Medium-term forecasting BdB Boaretto & Medeiros (2023) Use of individual prices from the statistical office for forecasting inflation up to 11 months ahead and testing the effectiveness of time series models and machine learning Use of individual prices leads to comparable results to models based on aggregate inflation. Machine learning methods are more efficient than traditional models
BIS Kohlscheen (2021) Random forests for inflation forecasting in 20 developed countries Random forests have a better predictive power than traditional OLS models
BoE Joseph et al. (2022) Using individual prices to predict headline inflation Non-linear machine learning methods like random forests are particularly effective in periods of significant changes in inflation
ECB Lenza et al. (2023) Quantile random forests Improved forecasts capturing non-linear inflation dynamics and forecast uncertainty
Fed Faria-e-Castro & Leibovici (2024) Use of genAI for US inflation forecasting LLM was more accurate than the Survey of Professional Forecasters
IMF Liu et al. (2024) Machine learning for post-pandemic core inflation forecasting in Japan Methods provide better inflation predictions, with LASSO being the most efficient

Note: BdB = Brazilian central bank, BIS = Bank for International Settlements, BoE = Bank of England, ECB = European Central Bank, Fed = Federal Reserve, IMF = International Monetary Fund, NBP = National Bank of Poland, OeNB = Austrian National Bank.

A2. Overview of AI Methods

Large language models (LLMs)

Large language models (LLMs) are modern artificial intelligence tools widely used across various fields, including inflation forecasting.

Notable models include OpenAI’s GPT (Generative Pre-trained Transformer), including variants such as GPT-4 and OpenAI o1. These models are widely implemented in applications like ChatGPT, where they are used for text generation, data analysis, translation and conversational management. In the context of inflation forecasting, they enable the processing and analysis of large volumes of economic data, including macroeconomic statistics, financial reports, media and social media content, and identify key trends that signal potential inflation pressures.

Another prominent model, Grok, comes from X.AI and is integrated with the X platform (formerly known as Twitter). This model is designed for real-time text analytics and focuses on sentiment analysis and trend detection in the social media environment. With these functionalities, it provides valuable insights into changes in consumer and business sentiment that can significantly affect inflation expectations.

Advanced models such as PaLM (Pathways Language Model) and Google’s Gemini introduce further innovations. PaLM is known for its deep text understanding capabilities and integration with Google products, while Gemini is designed for autonomous agents that manage complex tasks with minimal need for human intervention. These models can be used to analyse historical and current data such as price trends, geopolitical events and shifts in consumer demand, providing relevant predictions of economic indicators.

Microsoft’s Copilot is another example of a practical application of LLMs, particularly in tools such as Microsoft Office and Teams. This model enables the automation of administrative tasks, the generation of text documents, data analysis and scenario design based on a wide range of economic parameters. In the area of inflation, it can be used to quickly analyse data and generate detailed economic reports.

The Claude model from Anthropic stands out for its focus on safety and user-friendliness. At the CNB, Claude is being tested primarily for coding and algorithm development tasks that may accelerate the automation of key processes. The model allows users to work with sensitive economic data, both structured and unstructured, for example when analysing internal documents or price reports. These features can make it a useful tool not only for more accurate inflation predictions, but also for the development of applications supporting various analytical processes.

The RAG (Retrieval-Augmented Generation) method is another interesting technology being tested at the CNB in the area of document compliance with legal requirements. This method combines the text generation capabilities of language models with the ability to retrieve and verify relevant information from external data sources. The CNB is exploring its use for automated checking of prospectuses for bond issues, where it could replace time-consuming manual compliance checks.

Meta’s freely available LLaMA (Large Language Model Meta AI) model is particularly popular in academia. Its flexibility allows for testing alternative approaches to inflation prediction and conducting research of complex economic phenomena using historical data.

This list by no means covers all the possibilities of the current dynamic development of LLMs. Each model offers a unique approach to data analysis and inflation forecasting, and it is impossible to clearly predict which one will become the dominant tool in this field in the future.

Machine learning models

In addition to traditional economic approaches, machine learning methods are becoming increasingly important, opening up new possibilities for inflation prediction and economic data analysis. Basic methods, such as linear regression, allow for simple analysis of relationships between macroeconomic indicators, such as the impact of wages or unemployment on inflation. Logistic regression is suitable for modelling the probability of exceeding certain inflation targets.

Advanced methods include decision trees, such as random forests, which can analyse a large number of variables such as commodity prices, interest rates and consumer confidence. These methods are proving particularly useful in identifying key factors influencing short-term inflation pressures. Gradient boosting machines, such as XGBoost and LightGBM, provide higher predictive accuracy, which is important for short-term inflation estimates in turbulent periods, such as economic shocks or changes in monetary policy.

A special role among forecasting methods is played by the MIDAS (Mixed Data Sampling) model, which allows for the combination of data with different frequencies – for example, daily oil prices with monthly unemployment statistics. This method is particularly useful for inflation nowcasting, as it integrates the latest data with traditional macroeconomic indicators. At the CNB, MIDAS could be used to provide faster estimates of current inflation based on a wide range of data sources.

Another advanced method is SARMAX (Seasonal AutoRegressive Moving Average with eXogenous variables), which takes into account seasonal effects and external variables. This is crucial for predicting inflation in areas such as food and energy prices. At the CNB, SARMAX can be used, for example, to model the impact of seasonal factors on overall inflation.

Bayesian models, such as Naïve Bayes and Bayesian networks, help model probabilistic relationships between economic factors. Naïve Bayes is commonly used for quick classification of textual data, such as sentiment analysis in media reports, while Bayesian networks provide deeper insights into inflation cause-and-effect relationships.

Reinforcement learning, such as Q-learning, can be used to simulate different monetary policy scenarios and their effects on inflation. This approach allows central banks to optimise their decision-making in response to inflation pressures.

Other important approaches include AutoML and meta-learning, which automate model selection and optimisation, thereby speeding up the inflation forecasting process and allowing for more efficient use of available data. In addition, neural networks and deep learning techniques enable the analysis of complex time series, such as monitoring prices of goods and services and modelling macroeconomic indicators.

Dimensionality and clustering methods also play an important role. K-means clustering enables the grouping of regions or markets with similar inflation patterns, while principal component analysis (PCA) simplifies complex datasets and identifies key factors affecting inflation. These methods are actively used at the CNB in macroeconomic analyses and in the interpretation of modelling results.