Cryptoracle Data Analysis Team

This paper proposes an innovative approach that integrates Event Stream GPT (ESGPT) for continuous event modeling with a news-driven forecasting framework (News-to-Forecast, N2F) to systematically explore the predictive potential of CO indicators. By leveraging unified event representations, multimodal fusion, and interpretable prediction mechanisms, we construct a cross-modal forecasting architecture.

The CO indicators, which integrate price, trading volume, and community sentiment into a multidimensional metric system, are critical in the context of high-frequency trading within crypto markets. However, traditional approaches face two major challenges:

• Data heterogeneity: On-chain activities, community text, and news events are distributed across sparse and multimodal sources.

• Causal complexity: The transmission path—from KOL statements to sentiment shifts, capital flows, and ultimately price movements—is difficult to model explicitly.

To address these issues, this paper integrates two state-of-the-art technologies:

1) ESGPT: Encodes continuous-time event streams into a unified token sequence and captures intra-event causal dependencies via nested attention mechanisms.

2) N2F (News-to-Forecast) framework: Employs LLM agents to form a closed-loop pipeline of news filtering → impact classification → temporal forecasting → logical refinement.

Innovation: This is the first framework to combine ESGPT’s internal event modeling with N2F’s external news-driven reasoning, validated on a proprietary CO dataset from private crypto communitie

ESGPT introduces a novel modeling paradigm for continuous-time, multimodal event sequences with internal dependencies. It unifies heterogeneous and sparse events—such as on-chain interactions, community posts, and order book operations—into a standardized token sequence representation.Leveraging sparse storage and zero-shot evaluation techniques, ESGPT efficiently retains long-tail distribution patterns under constrained memory conditions, making it particularly well-suited for the high-frequency yet sparse nature of crypto market events.

In the CO setting, ESGPT captures internal drivers of indicator changes, including user behaviors, on-chain capital movements, and the triggering of parameterized trading strategies. It also provides attention weights and counterfactual generation capabilities, enabling downstream causal diagnostics and interpretability.

The N2F framework focuses on external information sources. It leverages large language models (LLMs) to build multi-turn agents that sequentially perform news retrieval, relevance filtering, short-/long-term impact classification, and reasoning feedback. High-confidence news events are then incorporated as exogenous variables into time series forecasting models. This framework has demonstrated effectiveness in tasks such as macroeconomic forecasting, electricity load prediction, and foreign exchange modeling, offering a feasible pathway to introduce macro-level and narrative-driven shocks into the crypto market prediction pipeline.

We adapt the event processing module originally designed for clinical time-series modeling to the Bitcoin market, aiming to predict market movements based on complex event sequences in continuous time. Specifically, given a sequence of heterogeneous events—such as social media discussions, KOL statements, and sentiment indicators—occurring over continuous timestamps, the objective is to model the following conditional probability distribution:

Given a sequence of historical community events, predict the occurrence time and characteristic performance of current market events (such as price fluctuations and trading volume changes).

And，

Include：

•  Categorical variables: event type, token symbol, community source

•  Continuous variables: sentiment score, propagation volume, capital flow

We will achieve this goal through a Transformer neural network architecture parameterized by θ, enabling the model to learn：

It is important to note that, unlike traditional GPT modeling scenarios, community data analysis often involves causal dependencies among feature variables within each event xi. For example, let xi(j) denote the j feature of the i community event (e.g., sentiment score, propagation volume, KOL influence), then:

Therefore, a complete generative model must fully account for the causal relationships among these internal features. For instance, fluctuations in community sentiment may directly impact propagation intensity, or the influence weight of KOL statements may modulate the effect of sentiment on price dynamics. The model explicitly captures these dependencies through a nested attention mechanism. 2.1.3 Data processing workflow

1. Multi-source Data Collection: Community texts from Telegram and Discord, on-chain transaction events.

2. Feature Engineering:

   Sentiment analysis using the VADER algorithm.

   KOL influence weight calculated as wKOL=log⁡(1+followers)×engagement rate

3. Sparse Storage: Memory usage scales linearly with the number of observed events

   2.2 News-to-Forecast Framework

The News-to-Forecast (N2F) framework proposes an innovative time series prediction method by iteratively inferring events from news text using large language models (LLMs) to achieve more accurate predictive performance. Its overall prediction process can be divided into four main modules: information acquisition, Agent inference and news screening classification, data integration and model fine-tuning, and closed-loop feedback optimization.

Specifically, the framework first acquires highly relevant news information and supplementary data (such as economic indicators, geographical, and meteorological information) from authoritative sources like GDELT and Yahoo Finance. Subsequently, it constructs an inference Agent driven by LLMs to screen news and determine their short-term or long-term impact to ensure the high relevance and accuracy of the input information.

To formally express this framework, we model news events as:

Among this, dj  is the description of the news text, cj  is the category of the news event judged by the LLM (short-term or long-term impact), and tj is the time when the news occurred.

After filtering by the Agent, the prediction task is transformed into:

Here, yt represents the historical time series observations, Nt denotes the news events filtered by the Agent, and θ are the model parameters.

By integrating both historical numerical sequences and textual news information, the model improves prediction accuracy and interpretability. In practice, the pretrained large language model is fine-tuned using LoRA (Low-Rank Adaptation) to further enhance the model’s joint understanding of numerical time series and news text data.

Coupling points between ESGPT and N2F:

Event embedding vectors output by ESGPT serve as temporal features for N2F.

News events filtered by N2F are fed back to optimize ESGPT’s event database training.

LoRA fine-tuning is applied to reduce GPU memory usage by 60%.

Attention heatmaps are used to visualize event impact pathways (see Figure 1).

1. Input: Existing CO indicator definition + community discussion text

2. Generate: LLM output for a new indicator proposal (name/formula/business value)

3. Evaluate: Automatically validate innovativeness and feasibility through Chain-of-Thought

Case: Generating the "Cross-Platform Sentiment Divergence" Indicator:

• Real-time Warning: TG group large transfer alert → Price fluctuation prediction

• Sentiment Transmission Analysis: KOL statement → Community sentiment diffusion → Trading volume surge

ESGPT provides a suite of innovative end-to-end solutions for researching the crypto market. Through a structured event flow modeling framework, it transforms fragmented market data and community activities into continuous time series event sequences. This technology overcomes the limitations of traditional analytical methods, enabling efficient processing of multimodal heterogeneous data, including price fluctuations, trading volume changes, social media sentiment, and KOL activities, among other complex information. Its core advantage lies in utilizing the Transformer architecture and nested attention mechanisms, which not only capture the complex interdependencies of market dynamics but also explicitly learn causal dependencies within events, such as the "sentiment-propagation-price" transmission path.

At the practical application level, ESGPT demonstrates strong generative modeling capabilities, capable of autoregressively generating future event sequences to support market forecasting, strategy backtesting, and anomaly detection. Leveraging the zero-shot transfer characteristics of pre-trained models, researchers can quickly adapt it to new crypto or community platforms. Additionally, the modular design allows for flexible expansion of data sources and adjustment of dependency graphs, while the visualization of attention weights enhances the model's interpretability, aiding in the identification of key market influencing factors.

The N2F framework provides a novel predictive paradigm for the quantitative analysis of cryptos. By leveraging the text generation and conditional probability prediction capabilities of large language models (LLMs), this framework innovatively integrates news events into time series predictions. It efficiently screens and categorizes news through iterative reasoning agents, enabling precise modeling of both short-term and long-term impacts of news events. Experimental validation with actual data, such as Bitcoin prices, demonstrates that the news-driven prediction framework significantly enhances prediction accuracy, especially during market unexpected event. Additionally, through a closed-loop feedback optimization mechanism, the model continuously refines its news screening logic, significantly reducing prediction errors introduced by irrelevant information, showcasing excellent adaptability.

•  Nestor, B., Chen, Y., & Caruana, R. (2024).  Event Stream GPT: A Data Pre-processing and Modeling Library for Generative, Pre- trained Transformers over Continuous-time Sequences of Com- plex Events.

•  Shen, Y., Wang, C., et al. (2024). From News to Forecast: Iterative Event Reasoning in LLM-Based Time Series Forecasting.