JoyfulFlame Co. Ltd(乐焰国际有限责任公司); Infinite Scenery Fund(无限风光基金）

In the crypto asset market, information asymmetry and uneven liquidity distribution are key drivers of sharp price volatility. Traditional price–volume factor models can partially explain routine market fluctuations, but they often fall short when facing extreme conditions triggered by external shocks, news-driven events, or abnormal capital flows.

Unlike conventional price–volume factors that rely solely on public market data, the core idea of the Cryptoracle factor is to uncover latent connections between private-domain social networks, KOL influence, and capital flows. By consolidating and quantifying fragmented social signals, Cryptoracle seeks to detect shifts in capital movements earlier and more sharply, transforming them into institution-grade trading intelligence.

The objective of this report is to systematically evaluate the predictive power of the Cryptoracle factor and its practical value at the trading level. Beyond assessing its effectiveness in forecasting returns and market direction, we place particular emphasis on its role in capturing abnormal volatility, enhancing risk management, and complementing traditional price–volume factors. This provides empirical evidence for factor research and the optimization of multi-factor models.

2.1.1 Raw Data Sources and Preprocessing

We utilize crypto asset candlestick (K-line) data—including open, high, low, close prices, trading volume, and turnover value—to compute traditional price–volume factors and generate return series. The data is sourced directly from the Binance API. To ensure accuracy in factor calculation and model training, the data is aligned by timestamp, deduplicated, cleaned for missing values, and normalized.

Sample Period: January 1, 2025 – August 14, 2025.

Training set: January 1, 2025 – May 14, 2025 (60%)

Testing set: May 15, 2025 – August 14, 2025 (40%)

Trading Pair Selection: Cryptoracle provided a factor library covering 200 cryptocurrencies. After excluding stablecoins (4) and tokens either not listed on Binance perpetual contracts or with incomplete data (9), a total of 187 trading pairs were retained (see Appendix for details).

2.1.2 Factor Construction

Cryptoracle Factors: Provided directly by our partner Cryptoracle. This study selects 13 factors: 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. (Detailed descriptions are available in the Cryptoracle official documentation.)

JF Factors: Developed independently by Joyful Flame Co., Ltd., forming a proprietary price–volume factor library currently deployed across several private funds, including the Infinite Scenery Fund.

Combined Factors (Both): An integrated factor library constructed by weighting both Cryptoracle and JF factors.

We evaluate the factors at two time resolutions: daily (1d) and 4-hour (4h). Due to the inherent characteristics of Cryptoracle factor collection, factor variance tends to be low and noise relatively high in shorter horizons such as 1h or 15min intervals. To ensure predictive validity and strategy robustness, we therefore restrict our analysis to the 4h and 1d frequencies.

2.1.3 Predictive Model

We employ an XGBoost model for multi-factor prediction (detailed parameters are provided in the Appendix). The model input is the factor feature matrix, and the output is the predicted return for the next time period.

Based on different factor combinations, we generate return forecasts on the test set. According to the prediction results, two types of trading strategies are designed to generate signals: beta-neutral and threshold-based.

2.2.1 Beta-Neutral Strategy

Using the predicted returns, the strategy goes long on the top n cryptocurrency trading pairs with the highest forecasted returns, and simultaneously goes short on the bottom n pairs with the lowest forecasted returns. Long and short positions are sized to maintain equal capital allocation, ensuring portfolio beta neutrality and reducing the impact of overall market volatility. Capital is distributed evenly across each selected long and short asset.

2.2.2 Threshold Filter Strategy

Given a predefined return threshold (threshold), the strategy goes long on all trading pairs with predicted returns greater than the threshold, and goes short on all trading pairs with predicted returns lower than –threshold. Capital is allocated equally across each selected long and short position.

Unlike the beta-neutral strategy, this approach does not fix the number of long/short assets. Instead, it dynamically selects trading pairs with stronger signals. By adjusting the threshold, the strategy can control its aggressiveness—balancing potential returns while limiting excessive diversification and noise trades.

Table 1: Evaluation of Predictive Performance of Cryptoracle Factors vs. JF Factors (MSE and Accuracy)

3.1.1 Daily (1d) Time Horizon

Using only the Cryptoracle factor library results in slightly higher training and test errors compared with the JF price–volume factor library, with accuracy close to 50%, indicating limited predictive power.

Using only the JF price–volume factor library yields lower MSE and slightly higher test accuracy than the Cryptoracle-only library (51.07% vs. 50.01%), showing stronger performance at the daily frequency.

The combined factor set (Both) produces test MSE values between the two single-factor models, with accuracy of around 50.76%, offering no significant improvement over individual factors.

3.1.2 4-Hour (4h) Time Horizon

At the 4h frequency, test MSE values increase significantly across all models, suggesting weaker predictive stability in higher-frequency settings.

The Cryptoracle-only factor library shows the largest test error (0.0045) and accuracy below 50% (49.14%), clearly indicating poor performance in high-frequency scenarios. This reflects the low variance and limited informational content of Cryptoracle factors at shorter horizons.

The JF factor library maintains relatively lower test MSE (0.0034) and accuracy around 50%, suggesting relatively stable performance at higher frequency.

The combined factor set (Both) performs similarly to the Cryptoracle-only case, with accuracy near 50% and no meaningful reduction in error, showing no evident advantage in high-frequency settings.

The predictive results show that daily data outperforms 4-hour data in various predictive aspects.

Furthermore, from the perspective of actual trading, daily trading can effectively reduce position management costs. Therefore, we only used daily data for backtesting on the test set.

3.2.1 Beta-Neutral Strategy

We tested the returns under different position distribution ranges (n values), as shown in Table 2.

Table 2 reports the backtesting performance of three factor-based strategies (Cryptoracle_only, JF_only, and Both) across different values of N (portfolio breadth). The following patterns can be observed:

(1) Narrow Portfolio Breadth (N = 1–10)

• When N is small, the strategy concentrates on a few trading pairs with the highest and lowest predicted returns.

• The Cryptoracle-only strategy exhibits relatively large return volatility. Positive returns are achieved at N=1 and N=10, but Sharpe ratios remain low, indicating higher risk.

• The JF-only strategy delivers significantly higher returns than Cryptoracle, with Sharpe ratios in the 4–5 range, reflecting robust risk-adjusted performance.

• The combined factor strategy (Both) achieves the highest overall returns, with Sharpe and Calmar ratios clearly outperforming single-factor strategies, demonstrating that factor combination enhances the risk–return profile in concentrated portfolios.

(2) Medium Portfolio Breadth (N = 20–30)

• As portfolio breadth expands, diversification increases and the contribution of individual assets declines.

• The Cryptoracle-only strategy shows negative returns at N=20 with a marked increase in maximum drawdown, suggesting weakened predictive power under diversification.

• The JF-only and Both strategies maintain positive returns, though the risk-adjusted performance of the combined strategy declines slightly.

(3) Broad Portfolio Breadth (N = 50)

• At highly diversified levels, returns decrease substantially.

• Both the Cryptoracle-only and combined strategies show significantly lower total returns and Sharpe ratios, while maximum drawdowns remain high.

(4) Overall Insights

• The Cryptoracle-only strategy performs better under concentrated allocation (small N), but its predictive power diminishes as diversification increases.

• The JF-only strategy demonstrates stable return distribution and risk control, making it suitable for medium- to large-scale portfolio construction.

• The combined strategy (Both) significantly outperforms in concentrated portfolios, with superior returns and risk-adjusted metrics, but suffers from return decay under broad diversification. This highlights the need to calibrate portfolio construction according to capital size and degree of diversification (see Figure 1).

3.2.2 Threshold filtering strategy

Table 3 shows that when using only Cryptoracle factors, total and annualized returns are negative or close to zero across almost all thresholds. Sharpe ratios are low or even negative, and maximum drawdowns exceed –200%, indicating excessive risk exposure without effective compensation. This suggests that relying solely on Cryptoracle factors under the threshold filter strategy lacks trading value.

Across all thresholds, using only JF factors yields Sharpe ratios in the range of 1.2–1.5,reflecting relatively stable risk-adjusted returns. However, maximum drawdowns remain severe (around –150%). When using the combined factor set, performance is not superior to JF-only at lower thresholds (0.1%–1%).As the threshold rises to 2%–3%, however, the combined strategy begins to show certain advantages, with both returns and sharpe ratios improving (e.g., at a 2% threshold, Sharpe = 1.04, higher than JF-only). This suggests that at lower thresholds (0.1%–1%), excessive noise dilutes predictive value, while as the threshold increases (around 2%), Cryptoracle factors begin to contribute marginal value in capturing extreme market movements.

(1) Limited effectiveness of Cryptoracle factors in standalone application

When applied independently for prediction and trading, Cryptoracle factors deliver limited overall returns and fail to generate a persistent trading edge. They are not recommended as standalone trading signals.

(2) Strong explanatory power for abnormal market movements

Cryptoracle factors demonstrate strong explanatory and detection capability for assets experiencing abnormal volatility. However, in more typical low-fluctuation scenarios, their explanatory power is weaker. In other words, these factors are better suited to capturing rare but significant market swings.

(3) Complementarity with traditional price–volume factors

Traditional price–volume factors are more effective in explaining routine market fluctuations but are less responsive to exogenous shocks or extreme market events. Cryptoracle factors provide an effective complement in this regard: when integrated into a price–volume factor model, they improve coverage across both normal and abnormal market conditions. As auxiliary signals within a multi-factor framework, they enhance performance in extreme environments, particularly by mitigating risks associated with abnormal volatility.

(4) Potential value in reducing drawdowns in multi-factor strategies

Within a multi-factor portfolio, incorporating Cryptoracle factors can significantly reduce drawdown risks caused by assets with abnormal volatility, thereby improving the overall robustness of the strategy.

(5) Further research potential

Cryptoracle factors demonstrate meaningful potential in abnormal market detection, risk control, and multi-factor portfolio optimization. They may serve as a key complementary component in the evolution of future multi-factor systems.

In this study, we adopted the XGBoost regression model (XGBRegressor) for factor-based forecasting. The parameter settings are as follows:

n_estimators = 5000

A relatively large number of boosting rounds was used to ensure sufficient learning capacity. This was paired with a lower learning rate (see below) to mitigate overfitting.

learning_rate = 0.01

A small learning rate was chosen so that each boosting step remains incremental, enhancing model stability and generalization.

max_depth = 6

The maximum depth of each tree was capped at 6, balancing model complexity and fitting power. This allows the model to capture nonlinear relationships while avoiding excessive overfitting from overly deep trees.

subsample = 0.8; colsample_bytree = 0.8

At each iteration, 80% of samples and features were randomly selected. This introduces randomness, improves generalization, and reduces the risk of overfitting.