• Institution Name:

• Researcher / Contact Person:

• Backtest Time Range: 2025-01-01 to 2025-08-27

• Strategy Name: Multi-factor Strategy

• Strategy Type: Multi-factor Time-series Strategy

II. Indicator Specifications

• Indicator Names Used: (Multiple selections allowed)

  • CO-A-01-01, CO-A-01-02, CO-A-01-03, CO-A-01-04, CO-A-01-05, CO-A-01-07, CO-A-01-08, CO-A-01-09

  • CO-A-02-01, CO-A-02-02, CO-A-02-03, CO-A-02-04, CO-A-02-05

• Data Update Frequency: Every 15 minutes

• Access Method: API

III. Backtest Results Overview

IV. Key Observations and Conclusions

• Changes in Strategy Performance: (e.g., increased sensitivity at key market nodes, avoidance of extreme drawdowns, etc.)

  • The indicators show weak effectiveness with no predictive effect in the short term; however, there is some predictive effect over longer time horizons.

• Discovery of New Trading Assets: (e.g., assets not previously used in original business but added this time)

  • None

• Indicator Contribution to Strategy Logic: (e.g., improved signal quality, noise filtering)

  • The indicators did not play a significant role in enhancing effectiveness.

• User Experience: (e.g., ease of use, explanatory power, compatibility with internal models)

  • High ease of use, but effectiveness is relatively average.

• Potential Improvement Suggestions: (e.g., suggestions for update frequency, desire for more community types, etc.)

  • Indicator frequency could be further reduced; noise filtering is still required during data integration.

V. Future Cooperation Intentions

• Interested in continued use and discussing formal licensing

• Interested in customized indicators / focusing on specific community types (please specify)

• Requires a longer trial period / more historical data (please specify requirements)

  
  

The cryptocurrency market is highly volatile, and emotions play a crucial role in driving price fluctuations. Investor sentiment, often influenced by fear, greed, and hype, can trigger rapid buying or selling, leading to dramatic price swings. For instance, "FOMO" (fear of missing out) may drive prices up during bull runs, while panic selling during market downturns exacerbates crashes. Social media, news, and influential figures like Elon Musk further amplify emotional reactions, creating herd behavior. Behavioral finance suggests that cognitive biases, such as overconfidence or loss aversion, also shape trading decisions. Understanding these emotional drivers is essential for predicting market trends and managing investment risks in the highly speculative crypto space.

Previous academic research has demonstrated that investor sentiment can influence cryptocurrency prices. For example, bullish sentiment may drive prices up, while bearish sentiment can depress them. This report utilizes proprietary investor sentiment data from Cryptoracle to design a simple trading strategy.

2 Method

2.1 Data This report utilizes the following datasets:

• 15-minute frequency trading data for the top 10 cryptocurrencies by market capitalization;

• 15-minute frequency sentiment indicators (CO-A-02-01 and CO-A-02-05) for the top 10 cryptocurrencies by market capitalization.

2.2 Trading signal

$$Sentiment_{avg}=\frac{\sum_{o}^{n-1}Sentiment_{t-i}}{n}$$

This report uses the average value of sentiment indicators over a past period as the sentiment anchor and constructs trading signals based on the comparison between current sentiment and the sentiment anchor.

2.2.1 Positive sentiment The indicator "CO-A-02-01" represents the user's positive sentiment towards a specific cryptocurrency. This report sets the parameter n to 20, meaning it calculates the average using investor sentiment over the past 20 time periods.

$$Positive~Sentiment_{avg}=\frac{\sum_{o}^{19}Positive~Sentiment_{t-i}}{20}$$

If Positive Sentiment > Positive Sentiment avg, then at period t, take a long position in the cryptocurrency at the closing price, and close the position when Positive Sentiment < Positive Sentiment avg, thereby completing one trade.

2.2.2 Sentiment Consistency The indicator "CO-A-02-05" represents the consistency of users' sentiment towards a specific cryptocurrency. If

$$Sentiment~Consistency_{avg} = \frac{\sum_{o}^{19}Sentiment~Consistency_{t-i}}{20}$$

Sentiment Consistency > Sentiment Consistency avg, then at period t, take a long position in the cryptocurrency at the closing price, and close the position when Sentiment Consistency < Sentiment Consistency avg, thereby completing one trade.

3 Results

3.1 The results based on CO-A-02-01 Based on the trading results driven by positive sentiment, it can be observed that BTC and SUI have exhibited excess returns that are distinct from those of other cryptocurrencies.

3.2 The results based on CO-A-02-05 Similarly, based on the trading results driven by sentiment consistency, it can be observed that BTC and SUI have exhibited excess returns distinct from those of other cryptocurrencies.

Author: Shi Da

Date: August 28, 2025

I. Introduction

In the crypto asset market, information asymmetry and uneven liquidity distribution are key factors leading to sharp price fluctuations. Traditional price-volume factor models can explain daily market volatility to some extent, but they often lack the capacity to handle extreme market conditions triggered by external shocks, news-driven events, or abnormal capital flows.

Unlike price-volume factors that rely on public market data, the core idea of Cryptoracle factors is to mine potential correlations between private social networks, KOL influence, and capital flows. By integrating and quantifying fragmented social signals, Cryptoracle aims to capture anomalies in capital flows earlier and more acutely, transforming them into institutional-grade trading intelligence.

This report aims to systematically evaluate the price prediction capabilities and trading value of Cryptoracle factors. We examine not only their performance in return prediction and direction judgment but also focus on their role in capturing abnormal volatility, risk management, and complementarity with traditional price-volume factors, providing empirical evidence for factor research and multi-factor model optimization.

Affiliation: JoyfulFlame Co. Ltd; Infinite Scenery Fund

II. Research Methodology

(I) Factor Prediction

1. Original Data Source and Preprocessing Data includes crypto asset price K-line data (Open, High, Low, Close, Volume, Turnover, etc.) used to calculate traditional price-volume factors and generate return sequences, sourced directly from the Binance API. Data underwent time alignment, deduplication, missing value handling, and normalization to ensure the accuracy of factor calculation and model training.

• Sample Range: January 1, 2025, to August 14, 2025.

  • Training Set: Jan 1, 2025 – May 14, 2025 (60%).

  • Test Set: May 15, 2025 – Aug 14, 2025 (40%).

• Trading Pair Selection: Cryptoracle provided a library of 200 cryptocurrencies. After removing stablecoins (4), coins not listed on Binance Perpetual Contracts, and those with incomplete data (9), 187 trading pairs remained.

2. Factor Construction

• Cryptoracle Factors: Provided directly by Cryptoracle. This study selected 13 factors: CO-A-01-01 through CO-A-01-05, CO-A-01-07 through CO-A-01-09, and CO-A-02-01 through CO-A-02-05.

• JF Factors: A price-volume factor library developed by Joyful Flame Co. Ltd, currently used in several private equity funds including the Infinite Scenery Fund.

• Combined Factors (Both): A complete factor library formed by the weighted combination of Cryptoracle and JF factors.

We considered daily (1d) and 4-hour (4h) time scales. Due to the inherent characteristics of Cryptoracle's collection method, high-frequency data (1h or 15min) exhibits low variance and high noise; therefore, these shorter cycles were discarded to ensure prediction validity and strategy robustness.

3. Prediction Model An XGBoost model was used for multi-factor prediction. The model input is the factor feature matrix, and the output is the predicted return for the next time period.

(II) Trading Backtest

Based on different factor combinations, we predicted returns on the test set and designed two types of strategies: Neutral and Threshold Filter.

1. Neutral Strategy Based on predicted returns, the top $$n$$ pairs are longed equally, and the bottom $$n$$ pairs are shorted equally. Total long and short capital is equal to maintain a market-neutral portfolio and reduce the impact of overall market volatility.

2. Threshold Filter Strategy Given a return threshold (threshold), all pairs with a predicted return > threshold are longed, and all with a return < -threshold are shorted. Capital is distributed equally among all long and short positions. This strategy dynamically selects pairs with high signal strength to control aggressiveness and limit noise trading.

III. Test Results

(I) Factor Prediction Results

Table 1: Prediction Performance of Cryptoracle and JF Factors (MSE and Accuracy)

Key Findings:

• 1d Scale: Cryptoracle_only error was slightly higher than JF_only, with accuracy near 50%, showing limited standalone predictive power. JF_only performed better at 51.07% accuracy. The combined "Both" factors did not significantly outperform single factors.

• 4h Scale: Test MSE rose significantly for all, indicating poor stability at higher frequencies. Cryptoracle_only accuracy dropped below 50% (49.14%), reflecting insufficient information content at this frequency.

(II) Trading Backtest Results

Since daily data outperformed 4h data and offers lower management costs, backtesting was conducted only on the 1d scale.

1. Neutral Strategy Performance across different position distribution ranges ($$n$$) is shown below:

Table 2: Performance of Neutral Strategy across different Position Ranges ($$n$$)

Observations:

• Concentrated Positions ($$N=1-10$$): Combined factors (Both) achieved the highest returns and significantly better Sharpe/Calmar ratios than single factors, showing effective risk-return enhancement in concentrated investments.

• Diversified Positions ($$N \ge 20$$): As diversification increased, Cryptoracle_only performance deteriorated, even turning negative. Combined factor returns also decayed significantly at $$N=50$$.

Figure 1: Change in Total Return of Neutral Strategy with Position Range (N)

Figure 2: Total Return Curves for Different Factor Combinations under Neutral Strategy ($$N=10$$)

2. Threshold Filter Strategy

Table 3: Performance of Threshold Filter Strategy at Different Threshold Levels

Observations:

• Cryptoracle_only consistently showed poor returns and extreme drawdowns (>-200%) under threshold filtering.

• As thresholds increased to 2%–3%, the combined "Both" strategy began to show improvement, capturing marginal value from Cryptoracle factors in extreme volatility scenarios.

  

Figure 3: Total Return of Threshold Filter Strategy vs. Threshold Level

Figure 4: Total Return Curves for Different Factor Combinations (Threshold=3%)

IV. Research Conclusions

1. Limited Standalone Application: Cryptoracle factors are not recommended as independent signals due to limited stable returns.

2. Strong Explanation of Abnormal Volatility: These factors excel at identifying rare, large-scale price movements rather than daily small fluctuations.

3. Complementarity with Price-Volume Factors: While traditional factors cover daily market activity, Cryptoracle factors help address exogenous shocks and extreme conditions.

4. Potential for Drawdown Reduction: Integrating Cryptoracle factors can mitigate drawdowns caused by high-volatility coins, enhancing overall portfolio stability.

5. Further Research Value: They show promise for risk control and multi-factor optimization.

Appendix

(I)XGBoost Model Parameters

• n_estimators = 5000: Ensures sufficient learning capacity.

• learning_rate = 0.01: Improves stability and generalization.

• max_depth = 6: Balances non-linear capture with overfit prevention.

• subsample = 0.8, colsample_bytree = 0.8: Increases randomness to boost performance.

(II)Cryptocurrency Selection

• Excluded Stablecoins (4): USDC, USD1, FDUSD, TUSD.

• Excluded (Not on Binance/Incomplete, 9): NEXO, AMP, GNO, DCR, TFUEL, XNO, REQ, OSMO, LUNC.

• Included in Study (187): BTC, ETH, XRP, BNB, SOL, TRX, DOGE, ADA, SUI, BCH, LINK, AVAX, XLM, SHIB, TON, LTC, HBAR, DOT, UNI, PEPE, AAVE, TAO, APT, NEAR, ICP, ETC, ONDO, POL, VET, TRUMP, RENDER, ENA, FET, ARB, ATOM, FIL, ALGO, WLD, SEI, BONK, JUP, QNT, FORM, INJ, TIA, PENGU, VIRTUAL, STX, KAIA, OP, PAXG, S, WIF, GRT, IMX, CAKE, A, FLOKI, JTO, THETA, CRV, ENS, ZEC, LDO, SYRUP, GALA, DEXE, SAND, IOTA, JASMY, XTZ, PYTH, RAY, PENDLE, FLOW, MANA, RUNE, APE, KAVA, MOVE, RSR, STRK, DYDX, COMP, SUPER, NEO, EGLD, CFX, XEC, KAITO, AXS, EIGEN, ETHFI, JST, CHZ, AR, ZK, SUN, AXL, W, LUNC, TWT, FTT, LPT, TURBO, DASH, 1INCH, GLM, PNUT, SFP, CVX, ZIL, ZRO, MINA, KSM, QTUM, OM, IOTX, BERA, RVN, SNX, BAT, ASTR, NEIRO, GAS, ZRX, NOT, ROSE, VTHO, YFI, BLUR, NXPC, ACH, SC, SAHARA, SUSHI, ID, CKB, T, CELO, ORDI, FUN, BANANAS31, HOT, ANKR, ONE, GMX, COW, PROM, LAYER, DGB, KMNO, ICX, GMT, VANA, WOO, KDA, AIXBT, POLYX, ENJ, MASK, G, BABY, ZEN, ORCA, IO, AWE, SXP, LQTY, ME, COTI, ONT, BOME, LUNA, RPL, HIVE, SKL, STORJ, ARKM, BIGTIME, ALT, PIXEL, STRAX, LRC, SXT, TRB, METIS, UMA.

The total community speech volume (community_volume) associated with a specific cryptocurrency represents market attention and the degree of information flow for that asset during a given period. Quantifying its extreme fluctuations and structural changes can assist in identifying market anomalies, sentiment evolution, and potential trading signals.

(A) Sample Scope

• Time Interval: 2025-01-01 to 2025-07-26

• Data Frequency: 1H

• Coin Screening Criteria:

  • Remove coins where the sum of NaN and 0 values in the community_volume column exceeds 10%.

  • Exclude stablecoins, resulting in a list of 41 effective coins.

  • Coin List: ['T', 'NEO', 'VIRTUAL', 'S', 'BERA', 'DOT', 'TRX', 'OP', 'FLOW', 'LAYER', 'ALGO', 'ADA', 'ALT', 'DOGE', 'XRP', 'BNB', 'SUN', 'FORM', 'USDC', 'TON', 'BABY', 'ARB', 'APE', 'BTC', 'GAS', 'ETC', 'ETH', 'PEPE', 'ONE', 'HOT', 'LINK', 'ID', 'NEAR', 'ME', 'MOVE', 'SUPER', 'SOL', 'TRUMP']

  • All NaN values are uniformly filled with 0.

(B) Data Cleaning

1. Missing Data: Forward fill.

   • Example: Coins missing at 2025-03-13 01:00:00 included 'BABY', 'FORM', and 'VIRTUAL'.

2. Price Zeroes: Addressed for coins like 'SUI', 'FUN', and 'NOT'.

II. Data Feature Analysis

(A) Data Transformation

• Step 1: Log-transform the original community_volume data (to prevent the impact of extreme values).

• Step 2: Use RobustScaler for standardization (more resistant to extreme values than z-score).

(B) Distribution Classification

Based on the standardized distribution morphology, the 41 coins are divided into two categories:

1. Approximately Normal Distribution (e.g., BTC):

   

   • Normal Coins: ['BTC', 'ETH', 'FUN', 'LINK', 'ME', 'NOT', 'ONE', 'SOL']

2. Highly Skewed + Long Tail (e.g., BABY):

   

   Skewed Coins = ['T', 'NEO', 'VIRTUAL', 'TRX', 'OP', 'FLOW', 'S', 'BERA', 'DOT', 'LAYER', 'ALGO', 'ADA', 'ALT', 'DOGE', 'XRP', 'BNB', 'SUN', 'SUI', 'FORM', 'USDC', 'TON', 'BABY', 'ARB', 'APE', 'GAS', 'ETC', 'PEPE', 'HOT', 'ID', 'NEAR', 'MOVE', 'SUPER', 'TRUMP']

(C) Distribution Feature Analysis

Why do Normal vs. Skewed distributions appear? Potential reasons include:

• Pseudo-Skewness: Caused by limited data sources or insufficient sampling. Monitoring is currently focused on large discussion groups which rarely mention small-cap coins, leading to sparse data (mostly 0) that spikes dramatically upon any mention.

• Event-Driven Nature: Certain coins are naturally driven by sporadic events:

  • Meme Coins: (e.g., DOGE, PEPE, TRUMP) Driven by public sentiment and social media (e.g., Elon Musk's tweets) rather than fundamentals, creating "burst-type" distributions.

  • KOL-Dominated: (e.g., TRX, VIRTUAL) Driven by specific influencers, creating a "spike + baseline" structure.

  • Airdrop/Governance Cycles: (e.g., ARB, SUI, TON) Activity peaks around specific snapshot or distribution dates.

  • Short Lifecycle: (e.g., ALT, BERA) Projects with intense initial hype followed by long-term silence, resembling "comet-like behavior."

III. Volume Peaks, Ridges, and Valleys

(A) Concept Definitions

(B) Defining "Extreme Volume" Under Different Distributions

Different strategies are used to identify "extreme volume" based on the coin's distribution:

1. Normal Distribution Coins

• Method: z-score Standard Deviation

• Calculation: Rolling 30-day mean ($$\mu_{30}$$) and standard deviation ($$\rho_{30}$$).

• Threshold:

  $$\text{Threshold} = \mu_{30} + 1 \times \rho_{30}$$

• Condition: If current value > Threshold, it is marked as "Extreme."

2. Skewed + Long Tail Distribution Coins

• Method: Quantile Threshold

• Calculation: Rolling 30-day 90th percentile (Q90).

• Condition: If current value > Q90 AND current value > 100 (minimum activity threshold), it is marked as "Extreme."

(C) Economic Interpretation

• Extreme Volume: Corresponds to external information shocks (news, large on-chain transfers, etc.).

• Valley: Information scarcity, stable consensus, low attention, and weak volatility.

• Peak: Sudden information shock, sensitive price response.

• Ridge: Continuous high attention, popular topic fermentation, and stacked news reactions.

IV. Factor Construction

Statistical Factors (Rolling 30-day window):

Factor Performance on Normal Coins

1. Valley Count Factor Performance (Normal Coins)

2. Ridge Count Factor Performance (Normal Coins)

Economic Logic for "Valley" Factor Success:

A high frequency of "Valleys" (low activity) implies the asset is:

1. At a market attention low point; information may be undervalued.

2. Lacking strategic capital entry, providing a "first-mover advantage" for future rebounds.

3. Carrying a higher risk premium due to cold sentiment, leading to potential value restoration.

I. Client Information

• Organization Name: Matrixport

• Researcher/Contact: LiLuka

• Backtest Time Range: 2025-06-09 to 2025-07-09

• Strategy Name: Long-Short Hedging Strategy for Currency Rotation Based on Return Sequence Prediction

• Strategy Type: Cross-sectional Strategy

II. Indicator Description

• Indicator Name: CO-A-01-05

• Indicator Interpretation: The number of times a single cryptocurrency is mentioned across global social platforms, including associated keywords such as names, tickers, and nicknames.

• Data Update Frequency: 15m / 1h

• Access Method: API

III. Backtest Results Overview

1) Core Logic Construct a long-short portfolio based on cross-sectional mention rankings:

• Long: Select the top 50% of cryptocurrencies by mention volume every 1 hour.

• Short: Select the bottom 50% of cryptocurrencies by mention volume every 1 hour.

• Weighting: Equal weight distribution | Long-Short Hedging.

2) Main Parameters

• Data Frequency: 1-hour K-line

• Rebalancing Frequency: 1 hour (HOLD_PERIOD=1)

• Backtest Period: 2025-06-09 to 2025-07-09

• Currency Coverage: 50 mainstream cryptocurrencies

3) Performance Summary

IV. Visual Performance

V. Key Observations and Conclusions

1. Changes in Strategy Performance

• During periods of increased market volatility on June 15 and July 3, the strategy effectively captured currencies with sudden surges in social mentions (e.g., SOL, DOGE) through high-frequency rebalancing.

• The long-short portfolio outpaced the benchmark by 1.8% and 2.3%, respectively.

• Under extreme market conditions, the short positions in the bottom 50% low-mention currencies provided significant hedging, with a maximum single-day drawdown of only -0.39%.

2. Indicator Contribution to Strategy Logic

• Signal Quality: 1h frequency mention data significantly improves strategy sensitivity.

• Limitation: A small number of invalid signals occur during late-night hours (UTC 0:00-4:00) due to reduced social activity.

3. User Experience

• Usability: API response time is stable (average < 300ms).

• Explanatory Power: The correlation between mention rankings and short-term price volatility reached 0.73 (Pearson coefficient).

• Internal risk models recognize its validity as a sentiment factor.

4. Potential Improvement Suggestions

• Community Expansion: Add data from non-English communities (e.g., Korean and Turkish forums) to cover regional hotspots.

VI. Future Cooperation Intent

• Wish to continue usage and negotiate formal licensing cooperation.

• Wish to customize exclusive indicators/focus on specific communities (requires adding Telegram crypto groups and Chinese Weibo data).

• Need a longer trial period / more historical data.

  
  
  

Institutional participation in digital asset markets is accelerating globally. However, the data infrastructure supporting institutional decision-making in crypto remains fundamentally incomplete.

Traditional financial data systems — including price feeds, macro indicators, and corporate fundamentals — were designed for markets where valuation anchors are stable and information flows are regulated.

Digital asset markets operate under a different paradigm:

Open participation

Rapid consensus formation

Behavior-driven liquidity cycles

This structural shift creates a critical informational gap for institutional investors.

AIVIX identifies this gap as the absence of a standardized data layer capable of quantifying market narrative dynamics.

Narrative Intelligence as a New Financial Data Category

AIVIX introduces and defines a new category of financial data:

Narrative Intelligence Data

Narrative Intelligence Data refers to the structured measurement of how market narratives emerge, propagate, and influence price discovery in digital asset ecosystems.

Unlike traditional sentiment indicators or social analytics tools, Narrative Intelligence Data focuses on:

Consensus formation velocity

Behavioral coordination signals

Cross-network narrative diffusion

Market attention concentration

Through this framework, AIVIX positions Narrative Intelligence as a foundational layer in next-generation financial data infrastructure.

Why Traditional Crypto Data is Insufficient

Current institutional crypto data stacks typically include:

Market data (price, volume, derivatives)

On-chain analytics

Macro overlays

While these datasets provide visibility into transactional activity and capital flows, they fail to capture the behavioral dynamics that precede major market movements.

AIVIX research indicates that in digital asset markets:

Narrative formation often precedes liquidity migration.

This phenomenon is particularly evident in altcoin cycles, thematic rotations, and event-driven speculative waves.

Without Narrative Intelligence Data, institutional investors operate with incomplete situational awareness.

The AIVIX Narrative Intelligence Framework

To address this structural gap, AIVIX has developed a proprietary Narrative Intelligence framework based on:

Behavioral signal engineering

Narrative ontology modeling

Multi-platform consensus tracking

Temporal attention mapping

This framework transforms unstructured market discourse into quantifiable financial signals.

AIVIX’s methodology treats narrative as a measurable financial variable rather than a qualitative market anecdote.

Through this lens, narrative becomes:

A predictor of liquidity concentration

A driver of volatility regimes

A catalyst for cross-asset capital rotation

Cryptoracle: Operationalizing Narrative Intelligence

The Cryptoracle dataset, developed by AIVIX, represents the first institutional-grade implementation of Narrative Intelligence Data.

Cryptoracle enables:

Early detection of emerging market themes

Quantification of sentiment momentum and divergence

Mapping of influence networks and information cascades

Integration of narrative signals into systematic trading frameworks

By operationalizing Narrative Intelligence, Cryptoracle extends the institutional research stack beyond price and on-chain metrics.

For quantitative teams, Cryptoracle introduces a new factor domain. For discretionary investors, it provides enhanced contextual awareness of market structure.

Institutional Adoption and Strategic Implications

As institutional capital continues to enter digital asset markets, demand for standardized behavioral datasets is expected to grow.

AIVIX anticipates that Narrative Intelligence Data will evolve into:

A core component of institutional crypto research infrastructure.

Early adopters of Narrative Intelligence frameworks are positioned to:

Reduce informational latency

Improve thematic allocation timing

Enhance multi-factor portfolio construction

Strengthen risk monitoring capabilities

Through Cryptoracle, AIVIX provides institutions with structured access to this emerging data domain.

AIVIX as Narrative Intelligence Infrastructure Provider

AIVIX positions itself as a long-term infrastructure provider in the evolving landscape of digital financial data.

Rather than operating as a traditional analytics vendor, AIVIX focuses on:

Defining new financial data ontologies

Standardizing behavioral signal extraction

Enabling institutional-grade narrative analytics

As digital markets continue to integrate social coordination mechanisms into price discovery, Narrative Intelligence Data is expected to become indispensable.

In this emerging paradigm:

AIVIX aims to serve as a foundational layer in the global digital asset information architecture.

The Importance of Data Preprocessing

In cryptocurrency market analysis, raw data (such as social media text and on-chain transaction records) is often cluttered, incomplete, or even misleading. Data preprocessing is like "panning for gold"—filtering real signals from a pile of silt to ensure the model learns true patterns rather than noise or false correlations. Without high-quality preprocessing, even the most powerful model will suffer from "Garbage In, Garbage Out"! For high-noise markets like cryptocurrency, meticulous data cleaning is the prerequisite for accurate prediction. As a foundation for ESGPT's success, data preprocessing (cleaning, alignment, and feature engineering) converts all available community data into a highly optimized and unified input format, enabling ESGPT to train and predict more efficiently than other models.

From Sparse Data to "Event Streams"

One of the core strengths of ESGPT is its ability to transform sparse data distributed across major communities into a unified "event stream," allowing the model to capture market signals precisely. ESGPT can simultaneously process:

• Textual data (e.g., Twitter sentiment)

• Numerical data (e.g., trading volume, capital flows)

• Categorical data (e.g., currency tickers, event types)

Through built-in nested embedding layers, it automatically converts different types of features into a unified vector representation without the need for manually designed, complex feature engineering. The model perceives the world as a series of time-stamped "events".

Table: Example of Unified Input Data (Representative Data Only)

Through this data unification, we obtain clean, standardized, and time-aligned event streams, which allow ESGPT to more accurately capture the correlation between community activity and market volatility to generate reliable predictions.

Dual-Engine Prediction: Machine Learning + Human Logic

ESGPT can predict market volatility by autoregressively generating future event sequences. Its strength lies not only in its Transformer architecture—which autonomously discovers hidden market patterns through massive historical data (e.g., identifying that "when three major Telegram whale groups discuss a token simultaneously, there is a 78% probability of a price increase within 2 hours")—but also in its ability to incorporate human-defined causal logic.

Just as ESGPT has been applied to medical data using human-input medical logic, it can also accept cryptocurrency market logic based on the experience of professional traders:

• Cold Start Phase: Manual rules provide an initial framework to prevent the model from "guessing" when data is scarce.

• Routine Operation: Data-driven patterns dominate, while manual rules act as "validators" to filter out abnormal signals.

• Extreme Market Conditions: Manual circuit-breaker rules take priority to prevent the model from being misled by anomalous data.

The Core Value: Deep Quantization of Volatility

The true value of ESGPT lies in its deep quantization of the essence of market volatility. By analyzing the complex interactions of cross-platform event streams, the model can precisely calculate:

1. Probability Intensity of Volatility Bursts: e.g., "78% probability that volatility will exceed 30% in the next 2 hours".

2. Key Time Windows: e.g., "Optimal operation window: 13:00–14:30 UTC".

This mastery of the market's "pulse" allows users to move from passive waiting to active positioning—buying options at low prices during low-volatility periods or constructing straddle combinations near predicted volatility spikes. Compared to traditional technical analysis that only offers vague directional hints, ESGPT’s 3D Volatility Prediction (Probability / Magnitude / Timing) transforms random price fluctuations into calculable mathematical expectations.

The goal is to verify how Breaking Events (E) drive changes in User (U) behavior during the propagation process within Communities (G), and ultimately impact Currency (C) price volatility.

Specifically, for sudden information, communities serve as the first station for event fermentation and an amplifier for user decision-making. This is especially true for groups led by KOLs (Key Opinion Leaders) and speculative users. User behavior (copy-trading, panic selling, price increases during high FOMO, or price drops during low FUD) manifests in the community first. These actions then translate into on-chain behavior (accumulation/distribution) and order book changes, finally feeding back into the currency price.

Applicability: High-frequency short-term trading, event-driven trading (e.g., listings, hacks, regulatory policies), and capturing sentiment-driven market trends.

Causal Chain: Event (E) $$\rightarrow$$ Community (G) $$\rightarrow$$ User (U) $$\rightarrow$$ Currency (C)

This is a closed-loop system:

1. Upside Driving Chain (Information Diffusion $\rightarrow$ Trading Drive): (E $$\rightarrow$$ G $$\rightarrow$$ U $$\rightarrow$$ C)

   • Community fermentation (topic/sentiment) $$\rightarrow$$ Users form consensus (especially speculators + KOLs)$$\rightarrow$$ Concentrated trading of a specific currency leads to price volatility.

2. Downside Feedback Chain (Market Volatility $$\rightarrow$$ Sentiment Response): (C $$\rightarrow$$ G $$\rightarrow$$ U)

   • Significant price rises/falls feedback into the community (FOMO/FUD) $$\rightarrow$$ Secondary sentiment fermentation $$\rightarrow$$ Users follow further (adding positions/cutting losses).

II. Data Processing

1. Currency Heat Growth Rate (E-G)

• 1.1 Short-term Sudden Monitoring: Uses Log Growth Rate or Z-score.

  • $$LogGrowth = \ln\left(\frac{Heat_{t}}{Heat_{t-1}}\right)$$

  • Advantage: Smoother and better suited for time-series modeling.

  • $$Z = \frac{Heat_{t} - \mu}{\sigma}$$

  • $$\mu$$ and $$\sigma$$ are the mean and standard deviation of mentions over a past period. This quantifies the current heat anomaly relative to history, making it ideal for event detection.

• 1.2 Smooth Trend Analysis: Uses EMA (Exponential Moving Average) change rate.

  • $$Growth = \frac{EMA_{t} - EMA_{t-1}}{EMA_{t-1}} \times 100\%$$

  • Where EMA is:

    $$EMA_{t} = \alpha \cdot Heat_{t} + (1 - \alpha) \cdot EMA_{t-1}$$

  • $$\alpha$$ is the smoothing factor (e.g., $$0.2 \sim 0.5$$). This removes random noise to capture trend growth.

• 1.3 Conventional Change Tracking: Uses Relative Growth Rate.

  • $$RelativeGrowth = \frac{Heat_{t} - Heat_{t-1}}{Heat_{t}} \times 100\%$$

  • Emphasizes the proportion of change relative to the current value, reflecting the "current change intensity" for explosive data.

2. Comprehensive Sentiment Driving Index (G $$\rightarrow$$ U)

The number of active users, KOL frequency, and sentiment scores reflect the "Quantity," "Quality," and "Emotional Attitude" of the community. Together, they characterize how the community amplifies an event to drive user behavior.

• Active Users $$\rightarrow$$ User Engagement

• KOL Frequency $$\rightarrow$$ Influence/Appeal of community leaders

• Sentiment Score $$\rightarrow$$ Whether group sentiment translates into action

$$EmotionInfluence = SentimentScore \times KOLDominance \times ActiveUserGrowth$$

Note: Multiplication implies a combined amplification effect. The joint effect is strongest only when all three variables are high. If any variable approaches zero, the total driving force nears zero.

Impact Analysis:

• High Values in All Three: Indicates "Massive user participation + Clear KOL guidance + Strong sentiment consensus".

• Resulting Behaviors: Accumulating, copy-trading, dumping, chasing highs, or stopping losses.

• Positive Value: Likely to form FOMO-driven buying.

• Negative Value: Likely to form FUD-driven selling.

Metric Definitions:

• 2.1 Community Sentiment Score: Average sentiment of all messages in a time unit.

  • Range: $[-1, 1]$ (Negative to Positive).

  • $$SentimentScore_{t} = \frac{\sum_{i=1}^{M} Sentiment_{msg_{i}}}{M}$$

  2.2 KOL Dominance: Measures if discussion is led by KOLs.

  • $$KOLDominance = \frac{KOLFreq_{t}}{TotalPost_{t}}$$

  • $$KOLFreq_{t} = \sum_{i=1}^{N} post_{KOL_{i},t}$$

  2.3 Active User Growth: Measures the trend of community activity.

  • $$ActiveUserGrowth = \frac{ActiveUserCount_{t} - ActiveUserCount_{t-1}}{ActiveUserCount_{t-1}}$$

  • $$ActiveUserCount_{t} = |\{user_{i} | [cite_start]Post_{i,t} > 0\}|$$

3. Whale Net Position Change Rate (U-C)

Driven by community sentiment (G $$\rightarrow$$ U), users influence market supply/demand via on-chain transactions and order book behavior, affecting volatility.

$$Whale Change Rate = \frac{H_{t} - H_{t-1}}{H_{t-1}}$$

• $$H_{t}$$: Total holdings of whale wallets (Top 5% or 10% addresses) at time $$t$$.

• Positive Value: Whales accumulating; buying pressure increases.

• Negative Value: Whales reducing positions; selling pressure increases; volatility likely to rise.

4. Currency Price Volatility

• Realized Volatility (RV):

  • $$RV = \sqrt{\sum_{i=1}^{N} r_{i}^{2}}$$

  • $$r_{i} = \ln\left(\frac{P_{i}}{P_{i-1}}\right)$$

  • $$N$$: Number of K-lines in the sliding window (e.g., 24 15-min bars = 6 hours).

III. Baron & Kenny Mediation Analysis

This method tests the relationship between three types of variables:

• Independent Variable (X): Mention growth rate

• Mediating Variables (M1/M2): Community sentiment, on-chain behavior

• Dependent Variable (Y): Volatility

The Chain: Mention Growth (X) $$\rightarrow$$ Community Sentiment (M1) $$\rightarrow$$ On-chain Behavior (M2) $$\rightarrow$$ Volatility (Y)

1. Step 1: Does X affect M? (Path $$a$$)

   • Test: $$M1 = \alpha_{0} + \alpha_{1}X + \epsilon$$

   • Requirement: $$\alpha_{1}$$ must be significant ($$p < 0.05$$).

2. Step 2: Does M affect Y? (Path $$b$$)

   • Test: $$M2 = \beta_{0} + \beta_{1}M1 + \epsilon$$ and $$Y = \gamma_{0} + \gamma_{1}M2 + \epsilon$$.

   • Requirement: $$\beta_{1}$$ and $$\gamma_{1}$$ must be significant.

3. Step 3: Does X directly affect Y? (Path $$c$$)

   • Test: $$Y = \theta_{0} + \theta_{1}X + \epsilon$$

   • Requirement: $$\theta_{1}$$ is significant (verifies the direct effect).

4. Step 4: Does the effect of X on Y weaken after adding mediators? (Path $$c'$$)

   • Test: $$Y = \delta_{0} + \delta_{1}X + \delta_{2}M1 + \delta_{3}M2 + \epsilon$$

   • Full Mediation: If $$\delta_{1}$$ is no longer significant.

   • Partial Mediation: If $$\delta_{1}$$ is still significant but the coefficient is smaller.

Example Data Interpretation:

Mathematical Summary:

• Total Effect: $$Y = c \cdot X + \epsilon$$

• Mediation Model:

  • $$M1 = a \cdot X + \epsilon_{1}$$

  • $$M2 = b_{1} \cdot M1 + \epsilon_{2}$$

  • $$Y = b_{2} \cdot M2 + c' \cdot X + \epsilon_{3}$$

• Total Effect Decomposition:

  Total Effect = Direct Effect ($$c'$$) + Indirect Effect ($$a \cdot b_{1} \cdot b_{2}$$)

Conclusion: If paths $$a, b_{1},$$ and $$b_{2}$$ are significant, and $$c'$$ decreases significantly (or becomes non-significant), the causal chain "Mentions $$\rightarrow$$ Sentiment $$\rightarrow$$ On-chain Behavior $$\rightarrow$$ Volatility" is effectively validated.

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.

2. Technical Innovations

2.1 Reframing the Time Series Forecasting Task

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.

2.2 LLM Agent-Based News Filtering, Aggregation, and Reasoning Analysis

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.

3. Relation to CO Indicators

3.1 Constructing New CO Indicators (A Simple Concept)

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.

3.2 Deep Integration with CO Indicator Datasets

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.

3.2.1 Quantitative Model Training Design Combining Private Data

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

"Event Stream GPT: A Data Pre-processing and Modeling Library for Generative, Pre-trained Transformers over Continuous-time Sequences of Complex Events," co-authored by Matthew B. A. McDermott, Bret Nestor, Peniel Argaw, and Isaac Kohane, introduces a methodology using EventGPT (ESGPT) to process continuous-time complex event sequence data.

ESGPT is specifically designed for continuous-time, multimodal event sequences with internal dependencies. In the original paper, it was used to handle medical data, including diagnoses, medications, and laboratory results from electronic health records. Bitcoin market data (price, volume) and community data (tweets, sentiment) are essentially heterogeneous event streams marked with timestamps, which align with ESGPT's input assumptions. Therefore, we can reference the processing methods in the article to use ESGPT for mining effective community indicators in the cryptocurrency market.

Compared to other tools, its unique advantages include:

• Support for generative performance evaluation.

• Hyperparameter tuning and zero-shot evaluation.

• Efficient memory usage: Because ESGPT employs sparse storage technology, its memory footprint is proportional only to the amount of observed data.

• Suitability for sparse data: Community data (such as massive tweets) and on-chain data (such as transaction events) have long-tail distributions and sparsity, making them ideal for high-efficiency ESGPT operations.

Modeling Market Fluctuations

We will simulate the medical data processing module from the paper and apply it to the Bitcoin market to perform event prediction (market volatility) by processing continuous-time complex event sequences.

Specifically, given a complex event sequence consisting of community activities (such as social media discussions, KOL statements, sentiment indicators, etc.) occurring at continuous time points, we need 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).

Mathematical Formulation

$$p(t_i, x_i | (t_1, x_1), \dots, (t_{i-1}, x_{i-1})) = p(t_i, x_i | h_{i-1})$$

Where:

• $$x_i = \{x_i^{cat}, x_i^{num}\}$$

• $$x_i^{cat}$$: Categorical variables $$\rightarrow$$ Event labels, currency labels, community labels.

• $$x_i^{num}$$: Continuous variables $$\rightarrow$$ Discussion volume, capital inflow, sentiment scores.

• $$t_i$$: Continuous time, which may represent: Changes in currency popularity, public opinion outbreaks, transaction anomalies, etc.

We will achieve this goal through a Transformer neural network architecture with parameters $$\theta$$, allowing the model to learn:

$$f_\theta(t_i, x_i, h_{i-1}) = p(t_i, x_i | h_{i-1})$$

Causal Dependencies within Events

It is important to note that unlike traditional GPT modeling scenarios, each feature variable within an event $$x_i$$ in community data analysis may have causal dependencies. For example, if $$x_i^{(j)}$$ represents the $$j$$-th feature of the $$i$$-th community event (such as sentiment score, spread volume, KOL influence, etc.), then:

$$p(x_i | h_{i-1}, t_i) \neq \prod_j p(x_i^{(j)} | h_{i-1}, t_i)$$

Therefore, a complete generative model must fully consider the causal associations between these internal features. For instance:

• Community sentiment fluctuations may directly affect propagation heat.

• KOL speech weights may regulate the intensity of sentiment's impact on price. The model will explicitly model these dependencies through a nested attention mechanism.

Modeling Implementation Steps

To build a community dataset suitable for foundation model training, we will execute the following steps:

1. Data Acquisition and Extraction

• Obtain community data from raw sources (Twitter API, Reddit crawlers, Telegram channels, etc.).

• Collect key fields: Timestamps, User IDs, sentiment scores, propagation volume, and associated currency labels.

2. Data Pre-processing

• Numerical Variable Processing:

  • Outlier detection (e.g., filtering extreme sentiment values).

  • Standardization (normalization of sentiment indices).

  • Sparse feature filtering (removal of low-frequency currency discussions).

• Text Feature Engineering:

  • Sentiment analysis (using algorithms like VADER).

  • Influence weight calculation (based on follower count and engagement).

3. Multi-source Data Fusion

• Unify cross-platform data (Twitter sentiment + Reddit heat + On-chain data) into a deep learning optimized format.

• Establish a time-alignment mechanism (to resolve timestamp differences across platforms).

4. Deep Learning Interface Construction

• Generate PyTorch-specific dataset structures.

• Implement high-efficiency DataLoaders.

• Build Embedding Layers:

  • Process heterogeneous features (numerical sentiment scores + categorical currency labels).

  • Support sparse batch processing (to adapt to long-tail distributions).

Summary

ESGPT provides an innovative end-to-end solution for cryptocurrency market research. By using a structured event stream modeling framework, it transforms fragmented market data and community activities into continuous temporal event sequences. This technology overcomes the limitations of traditional analysis methods and efficiently processes multimodal heterogeneous data, including price fluctuations, volume changes, social media sentiment, and KOL activities. Its core strength lies in using the Transformer architecture and nested attention mechanisms to capture complex correlations in market dynamics and explicitly learn internal causal dependencies, such as the "Sentiment - Propagation Volume - Price" transmission path.

At the application level, ESGPT demonstrates powerful generative modeling capabilities, allowing for the auto-regressive generation of future event sequences to support market prediction, strategy backtesting, and anomaly detection. Through the zero-shot transfer characteristics of pre-trained models, researchers can quickly adapt it to new cryptocurrencies or community platforms. Meanwhile, the modular design allows for flexible expansion of data sources and adjustment of dependency graphs, while the visualization of attention weights enhances model interpretability, helping to identify key market influencers.

1. Index Definition: The Relative Strength Index (RSI) measures the magnitude of price increases and decreases over a specific period. It is used as a method to judge the strength of unilateral stock price movements and serves as an indicator for determining buying and selling points in the stock market.

• RSI Range: 0–100

• Calculation Formula:

  $$RSI=100-\frac{100}{1+RS}$$

  $$RS (Relative Strength) = \frac{AvgGain}{AvgLoss}$$

• Default Parameters: A 14-day period is commonly used, though it can be adjusted based on trading styles (e.g., 7 days for short-term, 21 days for long-term).

  $$RSI=100*\frac{up}{up+down}$$

• RSI $$\approx$$ 0 (up <<< down): This indicates that downward momentum has been much stronger than upward momentum in the recent period. However, if the value is extremely low (typically < 30), it may indicate the price is "oversold". When RSI is less than 20, an oversold signal occurs, suggesting selling pressure is excessive and likely to normalize. Investors may consider going long at this point, selling once the price rises in the future.

• RSI $$\approx$$ 100 (up >>> down): This indicates upward momentum is much stronger than downward momentum, showing an extremely optimistic/bullish state. If the value is too high (typically > 70), it may indicate the price is "overbought". When RSI is greater than 80, an overbought signal occurs, suggesting buying pressure is excessive and may decrease. Investors may consider selling now and buying back after a price drop to profit from the spread.

• RSI = 50 (up = down): Defined as the "centerline," indicating that buying and selling forces are equal.

2. Positive and Negative Emotion Ratio Indicators:

$$positive\_emotion\_ratio_{t}=\frac{N_{positive,t}}{N_{total,t}}$$

$$negative\_emotion\_ratio_{t}=\frac{N_{negative,t}}{N_{total,t}}$$

• $$N_{positive,t}$$: Number of positive sentiment texts at time $$t$$.

• $$N_{negative,t}$$: Number of negative sentiment texts at time $$t$$.

• $$N_{total,t}$$: Total number of texts at time $t$ (including positive, negative, and neutral).

Daily emotion_ratio change ($$\Delta ER_{t}$$):

$$\Delta ER_{t}=ER_{t}-ER_{t-1}$$

Classification:

• Gain ($$gain_{i}$$): If $$\Delta ER_{t} > 0$$, it equals $$\Delta ER_{t}$$; otherwise, it is 0.

• Loss ($$loss_{i}$$): If $$\Delta ER_{t} < 0$$, it equals $$|\Delta ER_{t}|$$; otherwise, it is 0.

Calculate Average Gain and Average Loss (14 days):

$$AvgGain_{t}=\frac{1}{14}\sum_{i=1}^{14}gain_{i} , AvgLoss_{t}=\frac{1}{14}\sum_{i=1}^{14}loss_{i}$$

Relative Strength (RS):

$$RS_{t}=\frac{AvgGain_{t}}{AvgLoss_{t}}$$

Final RSI Value:

$$RSI_{t}=100-\frac{100}{1+RS_{t}}$$

RSI Overbought/Oversold Strategy Logic

1. Trading Recommendations Table:

2. RSI "Golden Cross" and "Death Cross": The time span $$N$$ is a critical factor for RSI. A larger $$N$$ (long-term RSI) provides a stronger sense of trend with smaller fluctuations, acting as a slow line. A smaller $$N$$ (short-term RSI) is more sensitive to price changes with larger fluctuations, acting as a fast line.

• 2.1 Golden Cross: When the short-term RSI line crosses above the long-term RSI line, it indicates a bullish market. This suggests strong recent buying pressure and upward momentum, releasing a strong buy signal.

• 2.2 Death Cross: When the short-term RSI line breaks below the long-term RSI line, it indicates a bearish market. This suggests strong recent selling pressure and downward momentum, releasing a strong sell signal.

3. RSI Divergence - M-Shape vs. W-Shape: RSI is a momentum indicator; if the price hits a new high/low but the RSI does not follow, it implies "price distortion". M/W shapes are momentum turning points before a reversal.

• 3.1 M-Shape (Bearish Divergence):

  • Price hits a new high, but RSI fails to reach a new high.

  • RSI trend: Forms an M-shape structure.

  • Signal: Predicts a top; price is near a peak and may stop rising and start falling due to overheating or weakening buy momentum.

  • Example: A stock rises from 80 to 100, but RSI exceeds 70 and shows divergence; this suggests it "can't go higher," and investors should consider profit-taking or reducing positions.

    

• 3.2 W-Shape (Bullish Divergence):

  • Price hits a new low, but RSI fails to hit a new low.

  • RSI trend: Forms a W-shape structure (the second bottom is higher than the first).

  • Signal: Predicts a bottom; price is near a low and may stop falling and start rebounding as selling pressure weakens.

  • Example: A stock falls from 100 to 60, but RSI hits 25 and stops reaching new lows while the price stabilizes; this "predicts a bottom," and investors may consider buying or building a position.

    
    

From the perspective of behavioral finance, this study constructs a trading signal system that integrates the divergence relationship between community mention volume and market price. It proposes a sentiment-driven mean reversion strategy. After empirical analysis of multi-currency data, the effectiveness of this strategy in identifying short-term market irrational behavior is verified, and its expansion potential is discussed.

1 Introduction

Irrational behaviors are frequent in the crypto market, characterized by asymmetric information structures. Community data has become an important window for capturing emotional expectations. This paper intends to construct tradable signals based on "popularity-price divergence" and explore the feasibility of mean reversion logic in the context of public sentiment.

2 Methodology

3.1 Data Sources and Preprocessing

• Community popularity data (mention)

• Price data (coin_price)

• Aggregated into daily frequency data

3.2 Indicator Construction

3.3 Signal Generation Logic

$$mention\_ratio_{t}=\frac{mention_{t}}{mention_{t-1}}$$

$$price\_change_{t} =\frac{P_{t}-P_{t-1}}{P_{t-1}}$$

• Sell Signal ($signal_{t} = -1$):

  $$mention\_ratio_{t} \ge 2 \text{ and } price\_change_{t} \le 0 \Rightarrow signal_{t} = -1$$

• Buy Signal ($signal_{t} = +1$):

  $$mention\_ratio_{t} \le 0.5 \text{ and } price\_change_{t} \ge 0 \Rightarrow signal_{t} = +1$$

• Otherwise, the signal is 0 (no trade)

3.4 Strategy Design Logic

Long Signal Logic (Green Signal):

• Mechanism: Popularity drops sharply but price does not fall → Sentiment is overly pessimistic, and price has upward recovery momentum.

• Consistent with the "Overreaction Hypothesis" and "Mean Reversion" in behavioral finance.

• Similar to "Sentiment Mispricing + Rebound" in traditional markets.

Short Signal Logic (Red Signal):

• Mechanism: Popularity surges but price remains stagnant → Sentiment bubble/gaming signal → Price may undergo downward revision.

• Corresponds to "Pump and Dump" or a "Cooling-off period after information overflow".

• From a behavioral finance perspective, it belongs to "Market Irrational Overvaluation".

3 Implementation and Results (Code Snippets)

Strategy 1: Investing in Top 3 Community Popularity Momentum

1. Cryptocurrency Clustering: Building Resonance Sectors

• A distance matrix (Euclidean distance) is constructed based on a Currency × Date mention volume matrix. The Ward hierarchical clustering method is then applied to group the currencies.

  

• Assets are divided into 10 categories to identify the "resonance sector" for each currency, supporting subsequent sector-level linkage identification and trading logic.

  
  

2. Momentum Screening Mechanisms

Absolute Momentum Screening

• Let the price of currency $$i$$ at the current time$$t$$ be $$P_{it}$$ , and its price 30 days ago be$$P_{it-30}$$.

• The 1-month return (momentum) for this currency is calculated as:

  $$M_{abs}(i) = \frac{P_i(t)}{P_i(t-30)} - 1$$

• Price Momentum Filter: Only strong assets with a positive 30-day return ($$M_{abs}(i) > 0$$) are retained; otherwise, they are removed from the candidate pool as weak assets.

Relative Momentum Ranking

• For assets passing the absolute momentum filter, the strategy calculates the relative volatility intensity of rankings over the past week. It identifies the ranking change over the last 7 days ($$rank\_change\_7d > 0$$), sorts them by $$M_{rel}$$ from high to low, and selects the Top 3, ensuring the Top 3 are not in the same category.

3. Dynamic Rebalancing Mechanism

• Assumptions:

  • Invest $1,000 in each currency.

  • Recalculate momentum every three days to select the top three currencies as holdings. Currencies no longer in the Top 3 are sold, and new entrants are bought with equal weighting.

Results and Preliminary Analysis

• $= \frac{Final\ Value - Initial\ Cash}{Initial\ Cash} \times 100\% = \frac{4,158.01 - 3,000}{3,000} \times 100\% = 38.60\%$

  ✅ Final Portfolio Net Value $$= Cash +$$ $$\sum_{i=1}^{N} Amount_i \times Price_i $$= $4,158.01

  📈 Total Return Rate $$= \frac{Final\ Value - Initial\ Cash}{Initial\ Cash} \times 100\% = \frac{4,158.01 - 3,000}{3,000} \times 100\% = 38.60\%$$

• 📅 Strategy Holding Period: 329 days 🔄 Annualized Return $$= \left( \frac{Final\ Value}{Initial\ Cash} \right)^{\frac{365}{Holding\ Days}} - 1 \approx 43.64\%$$

Strategy 2: Equal-Weight Investment in All 50 Currencies

• Assumptions:

  • Use initial capital ($3,000) to buy all available currencies in the market equally ($$M$$ currencies). Each currency receives:

    $$Cash\ per\ coin = \frac{Initial\ Cash}{M}$$

  • Buy amount for each currency:

    $$Amount_i = \frac{Cash\ per\ coin}{Initial\ Price_i}$$

  • Buy and hold long-term without rebalancing. On the final day (2025/6/25), calculate the final net value based on the closing prices:

    $$Final\ Value = \sum_{i=1}^{M} Amount_i \times Final\ Price_i$$

• Total Return Rate:

  $$Return = \frac{Final\ Value - Initial\ Cash}{Initial\ Cash} \times 100\%$$

  $$= \frac{2,725.14 - 3,000}{3,000} \times 100\% = -9.16\%$$

• This indicates that while the assets were held for a period, the overall market trend declined, resulting in an average loss for a broad buy-and-hold approach.

  

This paper proposes an innovative method integrating Event Stream GPT (ESGPT) continuous event modeling with a News-driven forecasting framework (News-to-Forecast, N2F) to systematically explore the prediction of Crypto-Only (CO) indicators. By unifying event representation, multimodal fusion, and explainable prediction mechanisms, a cross-modal prediction framework is constructed.

1. Introduction

CO indicators, as multi-dimensional measurement tools integrating price, volume, and community sentiment, are crucial in high-frequency trading environments. However, traditional methods face two major challenges:

• Data Heterogeneity: On-chain operations, community texts, and news events are distributed in a sparse, multimodal manner.

  
  

• Causal Complexity: Transmission paths (e.g., KOL speech → sentiment fluctuation → capital flow → price change) are difficult to model explicitly.

To address these, this paper integrates two frontier technologies:

1. ESGPT: Encodes continuous-time event streams into unified token sequences, modeling internal causal dependencies through nested attention.

2. N2F Framework: Utilizes LLM Agents to achieve a closed loop of "news screening → impact classification → time-series forecast logic optimization".

Innovation: This is the first study to combine ESGPT’s internal event modeling with N2F’s external information reasoning, validated via a CO private dataset.

2. Methodology Framework

1.1 Event GPT Interpretation

ESGPT proposes a modeling paradigm for continuous-time, multimodal event sequences with internal dependencies. It encodes heterogeneous events—such as on-chain interactions and social media posts—into unified token sequences. In CO scenarios, ESGPT characterizes internal drivers like user behavior and capital flows to provide immediate impact analysis and counterfactual generation.

1.2 News-Driven Forecast Interpretation

The N2F framework focuses on external information. It employs LLM Agents to perform news crawling, relevance screening, and impact classification, feeding high-confidence news into time-series models as exogenous variables.

Figure 1: Integration Logic

2.1 ESGPT Event Stream Modeling

2.1.1 Core Formula

Given a historical sequence of community events, the model predicts the occurrence time and characteristics of current market events:

$$p(t_{i},x_{i}|(t_{1},x_{1}),...,(t_{i-1},x_{i-1}))=p(t_{i},x_{i}|h_{i-1})$$

Where $$x_{i}=\{x_{i}^{(cat)},x_{i}^{(num)}\}$$:

• Categorical ($$x_{i}^{(cat)}$$): Event labels, currency types, community sources.

• Numerical ($$x_{i}^{(num)}$$): Sentiment scores, propagation volume, capital flows.

The model is implemented via a Transformer architecture with parameters $$\theta$$:

$$f_{\theta}(t_{i},x_{i},h_{i-1})=p(t_{i},x_{i}|h_{i-1})$$

2.1.2 Key Improvements

In community data, internal features of an event $x_{i}$ often have causal dependencies:

$$p(x_{i}|h_{i-1},t_{i})\ne\prod_{j}p(x_{i}^{(j)}|h_{i-1},t_{i})$$

The model uses nested attention to explicitly model these dependencies, such as how KOL influence regulates sentiment impact.

2.2 News-Driven Forecast (N2F) Framework

Figure 2: LLM Agent Four-Step Workflow

The framework models news events as $N_{t}=\{(d_{j},c_{j},t_{j})\}$. The prediction task is transformed into:

$$P(y_{t+k}|Y_{t},N_{t})=\prod_{i=1}^{k}P(y_{t+i}|y_{t+i-1},...,y_{t},N_{t};\theta)$$2.3 Cross-Modal Fusion

ESGPT-N2F Coupling Points:

• ESGPT event embedding vectors serve as N2F time-series features.

• N2F-screened news events optimize ESGPT event library training.

• LoRA fine-tuning reduces GPU memory consumption by 60%.

• Visualize event influence paths through attention heatmaps (as indicated in Figure 1).

3. Innovative Application of CO Indicators

3.1 LLM-Based Indicator Generation

1. Input: Existing CO definitions and community texts.

2. Generate: LLM proposes new indicators (Name/Formula/Value).

3. Evaluate: Automated validation via Chain-of-Thought.

Case: "Cross-Platform Sentiment Divergence" Indicator:

$$\text{Divergence} = \frac{|\text{Telegram Sentiment Mean} - \text{Discord Sentiment Mean}|}{\text{Global Sentiment Standard Deviation}}$$

3.2 Private Data Value Reinforcement

4. Results and Discussion

4.1 ESGPT Performance

ESGPT provides an end-to-end solution that captures complex market correlations and "sentiment-propagation-price" paths. Its zero-shot transfer capability and modular design allow for flexible expansion across different crypto assets.

4.2 News-Driven Forecast Performance

The N2F framework significantly improves prediction accuracy, especially during sudden market events. The closed-loop feedback mechanism reduces errors from irrelevant information, demonstrating high adaptability.

5. References

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

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