The time series forecasting method based on news event driving and Large Language Models (LLMs)—"From News to Forecast: Iterative Event Reasoning in LLM-Based Time Series Forecasting"—jointly proposed by Professor Zhao Junhua's team from the Chinese University of Hong Kong (Shenzhen) and Professor Qiu Jing's team from the University of Sydney, was recently accepted by NeurIPS, a top-tier AI conference. This paper introduces a new paradigm for time series forecasting: predicting time series data by combining LLMs with news texts. Training and validation on datasets such as power load, exchange rates, and Bitcoin prices demonstrate that this news-driven LLM approach outperforms existing time series forecasting methods, fully proving the potential of combining "alternative text information" like news in the field of cryptocurrency quantitative analysis.

The overall forecasting framework in the paper can be roughly divided into four steps: the Information Retrieval Module, constructing a Reasoning Agent to filter and classify news, Data Integration and Model Fine-tuning, and evaluating and optimizing the Agent based on prediction errors. The paper collects time-series-related news and supplementary information (weather, geography, economic indicators, etc.) from sources including the GDELT news database and Yahoo Finance. Utilizing the reasoning capabilities of the LLM, it filters news strongly related to the forecasting task, classifying them by short-term or long-term impact. Through multiple rounds of prompting, it gradually summarizes effective news filtering logic. The filtered news, supplementary information, and historical time series are then integrated into text prompts to fine-tune a pre-trained LLM, enabling it to predict future time series via conditional probability. Finally, the results are fed back to the Reasoning Agent to optimize the news filtering logic, forming a closed-loop iteration.

Traditional time series forecasting usually leverages the numerical characteristics of the series, using auto-regressive properties or introducing exogenous variables combined with numerical methods. In contrast, this paper innovatively treats time series as "digital token sequences," utilizing the LLM's text sequence generation capability for prediction. Essentially, a traditional LLM predicts the next word with the maximum conditional expectation based on the preceding text, which is analogous to time series forecasting.

Assume a time series exists as {"123", "456"}. Given the character sequence "123", the probability of predicting "456" can be expressed as an auto-regressive probability prediction process:

$$P("456"|"123") = P("4"|"123") P("5"|"4","123") P("6"|"45","123")$$

In LLMs, news events can be represented as a set of text tokens $\{ e_0, e_1, e_2 ... e_u \}$ to characterize events. LLMs use this news information as conditional input to perform predictions through the conditional probability $P(x_t | x_{0:t}, e_{0:u})$. Introducing $e_{0:u}$ provides critical context that influences the prediction of future values.

While LLMs possess some ability to generate time series predictions, performing few-shot forecasting by directly providing raw time series and news data remains difficult. First, controlling the output of time series is challenging because numerical tokens are less common. Second, the connection between news and time series usually needs to be derived from historical data, which exceeds the conventional scope of few-shot forecasting with LLMs.

The authors adopt a Supervised Fine-Tuning (SFT) method, training the LLM with paired time-series and news data formatted as text input-output pairs using the Low-Rank Adaptation (LoRA) method. Because inappropriate or unfiltered news can introduce noise—potentially degrading prediction performance—the process is paired with a Reasoning Agent and an Evaluation Agent to ensure data quality through iterative optimization.

The specific process is as follows: In the first iteration, the LLM builds news filtering logic based on the task domain and time; the Reasoning Agent filters news according to this logic, aligns it with the time series, and inputs it for initial model tuning. In each subsequent iteration, the model's predictions are validated against a randomly extracted validation set; the Evaluation Agent checks for missing news that impacted the prediction and feeds back to the Reasoning Agent to optimize filtering logic. This loop continues until the final iteration, where the Reasoning Agent integrates updates to generate the final news filter.

We can borrow the framework of "Event Filtering, Reasoning, Filtering + Prediction, Reflection, Improvement" from the literature. First, prepare definitions of existing CO indicators, business goals, community discussion texts, and typical cryptocurrency news events as prompts. Then, have the LLM generate new indicators—including names, definitions, suggested calculation formulas, and business value explanations. In this way, the LLM learns from existing CO indicators and designs new candidate features across dimensions like cross-platform comparisons, user stratification, and sentiment dynamics by combining unstructured user discussions with numerical features.

Next, utilize Chain-of-Thought (CoT) reasoning to automatically evaluate the innovation, interpretability, and computational feasibility of these new indicators, filtering out the most promising candidates. Finally, combine analyst feedback and historical data validation for multi-round iterative optimization to continuously improve the quality and business relevance of the generated CO indicators.

The literature achieved superior results in Bitcoin price prediction using LLMs and news text. Within this LLM + News + Time Series framework, the use of powerful text reasoning to achieve a closed loop of "Event Selection—Causal Reasoning—Prediction Generation" not only improves accuracy but also demonstrates how to systematically transform "external text events" into structured model inputs. This highly aligns with the underlying logic of our CO indicators. Especially in the "high uncertainty, strong sentiment-driven" cryptocurrency market, text signals can effectively fill the blind spots of traditional price and volume data.

The key to the method lies in providing valuable text data to the Reasoning Agent. The financial sentiment time-series database within the CO dataset, based on specific private data sources, ensures the quality and informativeness of the text provided. It saves customers time in collecting external texts while offering unique private data not available through public channels, bringing infinite possibilities to cryptocurrency time series prediction.

Following the prompt engineering style used for tuning agents in the literature, a structured prompt can be constructed combining the characteristics of the CO dataset:

Integrated Input:

{Historical BTC Price Data:
2024-06-30 06:00: $61,750
2024-06-30 07:00: $62,010
2024-06-30 08:00: $61,500
2024-06-30 09:00: $61,020
2024-06-30 10:00: $61,200

Supplementary Information:
Current Market Volatility: 75% annualized, USDT market cap remains stable.
Large On-chain Transfer: 2024-06-30 11:15, over 1 billion USDT transferred out.

News Summary and Rationality:
2024-06-30 10:30, Twitter KOL leaks that an exchange is suspected of bankruptcy, which may intensify market panic and trigger BTC selling pressure in the short term.
2024-06-30 11:15, On-chain data shows abnormal fund movement, possibly related to panic or hedging demand.

Forecasting Task:
Based on the above information, predict the BTC price trend from 2024-06-30 12:00 to 2024-06-30 18:00.}

http://googleusercontent.com/immersive_entry_chip/0
http://googleusercontent.com/immersive_entry_chip/1
http://googleusercontent.com/immersive_entry_chip/2
http://googleusercontent.com/immersive_entry_chip/3

The time series forecasting method based on news event driving and Large Language Models (LLMs)—"From News to Forecast: Iterative Event Reasoning in LLM-Based Time Series Forecasting"—jointly proposed by Professor Zhao Junhua's team from the Chinese University of Hong Kong (Shenzhen) and Professor Qiu Jing's team from the University of Sydney, was recently accepted by NeurIPS, a top-tier AI conference. This paper introduces a new paradigm for time series forecasting: predicting time series data by combining LLMs with news texts. Training and validation on datasets such as power load, exchange rates, and Bitcoin prices demonstrate that this news-driven LLM approach outperforms existing time series forecasting methods, fully proving the potential of combining "alternative text information" like news in the field of cryptocurrency quantitative analysis.

The overall forecasting framework in the paper can be roughly divided into four steps: the Information Retrieval Module, constructing a Reasoning Agent to filter and classify news, Data Integration and Model Fine-tuning, and evaluating and optimizing the Agent based on prediction errors. The paper collects time-series-related news and supplementary information (weather, geography, economic indicators, etc.) from sources including the GDELT news database and Yahoo Finance. Utilizing the reasoning capabilities of the LLM, it filters news strongly related to the forecasting task, classifying them by short-term or long-term impact. Through multiple rounds of prompting, it gradually summarizes effective news filtering logic. The filtered news, supplementary information, and historical time series are then integrated into text prompts to fine-tune a pre-trained LLM, enabling it to predict future time series via conditional probability. Finally, the results are fed back to the Reasoning Agent to optimize the news filtering logic, forming a closed-loop iteration.

Traditional time series forecasting usually leverages the numerical characteristics of the series, using auto-regressive properties or introducing exogenous variables combined with numerical methods. In contrast, this paper innovatively treats time series as "digital token sequences," utilizing the LLM's text sequence generation capability for prediction. Essentially, a traditional LLM predicts the next word with the maximum conditional expectation based on the preceding text, which is analogous to time series forecasting.

Assume a time series exists as {"123", "456"}. Given the character sequence "123", the probability of predicting "456" can be expressed as an auto-regressive probability prediction process:

$$P("456"|"123") = P("4"|"123") P("5"|"4","123") P("6"|"45","123")$$

In LLMs, news events can be represented as a set of text tokens $\{ e_0, e_1, e_2 ... e_u \}$ to characterize events. LLMs use this news information as conditional input to perform predictions through the conditional probability $P(x_t | x_{0:t}, e_{0:u})$. Introducing $e_{0:u}$ provides critical context that influences the prediction of future values.

While LLMs possess some ability to generate time series predictions, performing few-shot forecasting by directly providing raw time series and news data remains difficult. First, controlling the output of time series is challenging because numerical tokens are less common. Second, the connection between news and time series usually needs to be derived from historical data, which exceeds the conventional scope of few-shot forecasting with LLMs.

The authors adopt a Supervised Fine-Tuning (SFT) method, training the LLM with paired time-series and news data formatted as text input-output pairs using the Low-Rank Adaptation (LoRA) method. Because inappropriate or unfiltered news can introduce noise—potentially degrading prediction performance—the process is paired with a Reasoning Agent and an Evaluation Agent to ensure data quality through iterative optimization.

The specific process is as follows: In the first iteration, the LLM builds news filtering logic based on the task domain and time; the Reasoning Agent filters news according to this logic, aligns it with the time series, and inputs it for initial model tuning. In each subsequent iteration, the model's predictions are validated against a randomly extracted validation set; the Evaluation Agent checks for missing news that impacted the prediction and feeds back to the Reasoning Agent to optimize filtering logic. This loop continues until the final iteration, where the Reasoning Agent integrates updates to generate the final news filter.

We can borrow the framework of "Event Filtering, Reasoning, Filtering + Prediction, Reflection, Improvement" from the literature. First, prepare definitions of existing CO indicators, business goals, community discussion texts, and typical cryptocurrency news events as prompts. Then, have the LLM generate new indicators—including names, definitions, suggested calculation formulas, and business value explanations. In this way, the LLM learns from existing CO indicators and designs new candidate features across dimensions like cross-platform comparisons, user stratification, and sentiment dynamics by combining unstructured user discussions with numerical features.

Next, utilize Chain-of-Thought (CoT) reasoning to automatically evaluate the innovation, interpretability, and computational feasibility of these new indicators, filtering out the most promising candidates. Finally, combine analyst feedback and historical data validation for multi-round iterative optimization to continuously improve the quality and business relevance of the generated CO indicators.

The literature achieved superior results in Bitcoin price prediction using LLMs and news text. Within this LLM + News + Time Series framework, the use of powerful text reasoning to achieve a closed loop of "Event Selection—Causal Reasoning—Prediction Generation" not only improves accuracy but also demonstrates how to systematically transform "external text events" into structured model inputs. This highly aligns with the underlying logic of our CO indicators. Especially in the "high uncertainty, strong sentiment-driven" cryptocurrency market, text signals can effectively fill the blind spots of traditional price and volume data.

The key to the method lies in providing valuable text data to the Reasoning Agent. The financial sentiment time-series database within the CO dataset, based on specific private data sources, ensures the quality and informativeness of the text provided. It saves customers time in collecting external texts while offering unique private data not available through public channels, bringing infinite possibilities to cryptocurrency time series prediction.

Following the prompt engineering style used for tuning agents in the literature, a structured prompt can be constructed combining the characteristics of the CO dataset:

Integrated Input:

{Historical BTC Price Data:
2024-06-30 06:00: $61,750
2024-06-30 07:00: $62,010
2024-06-30 08:00: $61,500
2024-06-30 09:00: $61,020
2024-06-30 10:00: $61,200

Supplementary Information:
Current Market Volatility: 75% annualized, USDT market cap remains stable.
Large On-chain Transfer: 2024-06-30 11:15, over 1 billion USDT transferred out.

News Summary and Rationality:
2024-06-30 10:30, Twitter KOL leaks that an exchange is suspected of bankruptcy, which may intensify market panic and trigger BTC selling pressure in the short term.
2024-06-30 11:15, On-chain data shows abnormal fund movement, possibly related to panic or hedging demand.

Forecasting Task:
Based on the above information, predict the BTC price trend from 2024-06-30 12:00 to 2024-06-30 18:00.}

http://googleusercontent.com/immersive_entry_chip/0
http://googleusercontent.com/immersive_entry_chip/1
http://googleusercontent.com/immersive_entry_chip/2
http://googleusercontent.com/immersive_entry_chip/3