Author: Shi Da

Date: August 28, 2025

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

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.

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.

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.

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%)

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.

• 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.

• 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.

Author: Shi Da

Date: August 28, 2025

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

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.

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.

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.

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%)

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.

• 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.

• 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.