⚙ Factor Construction Method

This system uses a Six-Factor Quadrant Model (V6 upgrade) for cross-sectional analysis of sector ETFs. It standardizes six dimensions — RSRS resistance/support strength, capital flow (Flow), price momentum (Momentum), financial quality (F_Quality), intraday efficiency (IntEff), and RSI momentum (RSI_Mom) — then combines them into a composite factor using weighted averaging.

• RSRS Factor (Resistance Support Relative Strength): OLS regression of daily highs and lows (β × R²), measuring support/resistance structural stability. Positive values indicate stable trend structure; negative values indicate unstable structure.
• Flow Factor: EWMA-weighted regression slope on ETF share data (half-life 3 days), then cross-sectional Rank standardization. Positive = capital inflow; negative = outflow.
• Momentum Factor: Price return over the past M days, divided by the same-period volatility (60-day) for risk adjustment.
• Financial Quality Factor (F_Quality): Combines three fundamental dimensions — ROE, PB valuation percentile (inverted), and earnings growth — of industry constituent stocks, using circulating market-cap weighted synthesis. See financial quality factor section below.
• Intraday Efficiency Factor (IntEff): Calculates daily price efficiency from OHLC proxy (Eff = net daily change / intraday path length), then takes the long-short EWMA difference on the 5-day SMA of Eff. Measures momentum changes in efficiency levels: positive values indicate improving efficiency (trend strengthening), negative values indicate declining efficiency (trend weakening / transitioning to oscillation). See intraday efficiency factor section below.
• RSI Momentum Factor (RSI_Mom): Computes the difference between short-term RSI(5) and medium-term RSI(20), then applies size neutralization by subtracting 0.5 × rank(fd_share). The rank-based approach and subtraction neutralization avoid collinearity with existing factors (max |Spearman| < 0.22). Captures short-term mean-reversion signals relative to medium-term momentum: positive values indicate short-term strength relative to trend; negative values indicate short-term weakness. See RSI momentum factor section below.

Weights w_* vary by preset (see Preset Parameters below), and the financial quality weight floats between 0.10 and 0.40 based on market timing signals.

All factors are processed via Rank standardization (standard score ranking), which is better suited for the cross-section of 32 sector ETFs compared to conventional Winsorize + Z-Score: it avoids statistical distortion from truncating tail samples and is naturally robust to outliers.

RSRS and momentum have very low collinearity (Pearson < 0.23), providing completely independent informational value.

★ Financial Quality Factor (F_Quality) — Added in V4

V4 introduced a fourth factor dimension — Financial Quality Factor (F_Quality) — providing cross-sectional screening capability from the perspective of listed company financial fundamentals. This factor is synthesized from three sub-factors with equal weighting:

• Latest ROE (F_ROE): Circulating market-cap weighted average of the latest quarterly ROE of industry constituent stocks. Measures overall industry profitability.
• PB Valuation Percentile (F_PB_Pct): The percentile rank of the latest PB_TTM among industry constituent stocks within their 5-year history, taken as an inverse indicator (1 - percentile). Lower valuation (lower PB historical percentile) scores higher.
• Earnings Acceleration (F_Earnings_YoY): Industry average of constituent stocks' single-quarter net profit YoY growth rate, then cross-sectional Z-Score. Reflects industry earnings growth momentum.

(Note: In the V4 stage, the composite factor used equal-weighted four factors; V5 added IntEff and switched to differentiated weights across five factors.)

Industry Constituent Stock Mapping: Each sector ETF uses 10 representative constituent stocks as a proxy portfolio, prioritizing circulating market-cap weighting, downgrading to equal weighting when data is insufficient. When fewer than 3 valid constituents are available, the industry factor is assigned the cross-sectional median.

Data Source: ROE and net profit growth from Tushare fina_indicator; PB and circulating market cap from stock_daily_basic.

Commodity ETF Handling: Commodity ETFs such as oil ETF and gold ETF do not calculate financial factors. The underlying code implements asset masking + weight redistribution: during composite factor synthesis, the F_Quality weight of commodity ETFs is proportionally redistributed to the four technical factors RSRS/Flow/Mom/IntEff (de-weighting formula: factor = technical weighted sum / (1 - w_quality)), ensuring fair cross-sectional scoring under the same dimensions.

Fault Tolerance: If a constituent stock has missing data, it is excluded from that period's calculation; if an entire industry has fewer than 3 valid data points, the industry factor is assigned the cross-sectional median.

⇄ Intraday Efficiency Factor (IntEff) — Added in V5

V5 introduced a fifth factor dimension — Intraday Efficiency Factor (IntEff) — measuring trend smoothness from the perspective of price movement structure. Rather than using the absolute efficiency value directly, this factor captures the direction of efficiency changes through efficiency indicator MACD (long-short EWMA difference).

• Underlying Efficiency (Eff): Daily OHLC proxy indicator, calculating the ratio of net daily change (|Close(t)-Close(t-1)|) to total intraday path ((High-Low)+|Close-Open|). High Eff = strong one-sided trend with low noise; Low Eff = frequent intraday reversals.
• Signal Smoothing: Optional SMA smoothing on Eff (short-term preset sma=0 skips this step to reduce lag; medium/long-term presets sma=5 for 5-day moving average), eliminating single-day anomalies.
• Momentum Quantification: Two EWMAs (half-life 5 days and 20 days) are applied to the smoothed series; the difference is IntEff. Positive = efficiency improving (trend strengthening); negative = efficiency declining (trend weakening / transitioning to oscillation).

IntEff participates in composite factor synthesis after cross-sectional Rank standardization. The core value of this factor lies in identifying inflection points in trend structure: when IntEff turns from negative to positive, it signals declining noise and potential trend acceleration; when it turns from positive to negative, it signals trend exhaustion risk.

↕ RSI Momentum Factor (RSI_Mom) — Added in V6

V6 introduced a sixth factor dimension — RSI Momentum Factor (RSI_Mom) — capturing short-term mean-reversion signals through multi-period RSI divergence, with size neutralization to maintain independence from existing factors.

• RSI Divergence Signal: RSI(5) - RSI(20) measures the gap between short-term and medium-term relative strength. A positive gap indicates short-term strength relative to the trend; a negative gap indicates short-term exhaustion.
• Size Neutralization: Subtracting 0.5 × cross-sectional rank of ETF fund shares (fd_share) reduces the factor's correlation with the quality and momentum factors. Backtesting confirmed max |Spearman| < 0.22 with all existing factors.
• Rank-Based Processing: Cross-sectional rank normalization followed by Z-score standardization, robust for the small-N (17-22 ETFs) panel.

Backtest results (short preset, H=10): solo IC=0.042, ICIR=0.144. Adding to the composite improved overall IC from 0.143 to 0.160 and ICIR from 0.501 to 0.545, with no significant collinearity with existing factors.

■ Quadrant Classification Model

Based on the sign of Z_Flow and Z_Mom, ETFs are classified into four quadrants:

• Q1 Strong Z_Flow ≥ 0, Z_Mom ≥ 0 — Capital inflow + price rising, trend resonance. Suitable for holding or adding positions.
• Q2 Lurk Z_Flow ≥ 0, Z_Mom < 0 — Capital inflow but price declining. Left-side positioning opportunity.
• Q3 Exit Z_Flow < 0, Z_Mom < 0 — Capital outflow + price declining. Highest risk, avoid.
• Q4 Risk Z_Flow < 0, Z_Mom ≥ 0 — Capital outflow but price still rising. Beware of bull traps.

Investment Implications: The investment recommendation engine prioritizes screening candidates from Q1 (60% budget) and Q2 (40% budget). In Q3, instruments with RSRS factor > 0.3 and composite factor > 0 are included as candidates based on signal strength (RSRS override mechanism); Q4 is always excluded.

● IC Analysis

IC measures a factor's predictive power for future returns. Uses Spearman rank correlation coefficient (Rank IC):

Look-Ahead Bias Control: Factors can only be calculated after T-day close (requiring T-day closing price and T-day share data), so IC analysis uses T-day factors to predict returns starting from T+1, ensuring no look-ahead bias.

Key statistics:

• IC Mean: Average of all daily IC values. |IC Mean| > 0.03 is significant, indicating factor predictive power.
• ICIR: IC Mean / IC Standard Deviation. |ICIR| > 0.05 is weakly effective, > 0.10 is usable, > 0.20 is good. Note: cross-sectional ICIR within industries is typically much lower than in stock selection; |ICIR| consistently > 0.05 in daily rolling windows has practical value.
• IC Win Rate: Proportion of trading days with IC > 0. > 50% indicates correct factor direction, > 60% is excellent.

⚗ Chart Reading Guide

1. Factor Distribution Histogram

Composite factor distribution of all ETFs for the latest period. Positive skew indicates most sectors are strong; negative skew indicates weakness.

2. Quadrant Return Heatmap

Average forward returns (%) for the four quadrants on the latest date. Ideal state: Q1 > Q2 > Q3 > Q4.

3. IC Time Series Curve

Daily IC values over time, including IC mean line and ±2σ volatility band. Sustained positive IC indicates stable factor predictive power.

4. Latest Factor Rankings

All ETFs ranked by composite factor in descending order, with quadrant labels. Provides an intuitive view of the strongest and weakest sectors.

5. Group Cumulative Return Curves

Cumulative returns of the four quadrants over time. The larger the gap between Q1 and Q3, the better the factor's discriminating power.

6. Rolling Factor ICIR

ICIR calculated using a 60-day rolling window. Rolling ICIR consistently > 0.05 indicates the factor has practical predictive power during that period.

⚙ Preset Parameters

Each preset defines six-factor parameters and differentiated weights:

• Optimized (Recommended): RSRS=20, Flow=10, Mom=20, H=15, 32-ETF pool. Weights: RSRS 0.38 / Flow 0.22 / Mom 0.32 / Quality 0 / Efficiency 0 / RSI_Mom 0.08. Portfolio stickiness=1.0 (favors holding continuity). RSRS MA20 dampening=0.5.
• Short-term: RSRS=20, Flow=10, Mom=20, H=10. Weights: RSRS 0.258 / Flow 0.129 / Mom 0.258 / Quality 0.184 / Eff 0.092 / RSI_Mom 0.08.
• Medium-term: RSRS=20, Flow=20, Mom=60, H=20. Weights: RSRS 0.193 / Flow 0.193 / Mom 0.258 / Quality 0.184 / Eff 0.092 / RSI_Mom 0.08.
• Long-term: RSRS=30, Flow=40, Mom=120, H=40. Weights: RSRS 0.161 / Flow 0.161 / Mom 0.322 / Quality 0.184 / Eff 0.092 / RSI_Mom 0.08.

Optimized Preset Design: Quality (ICIR=-0.24) and Efficiency (ICIR=-0.19) were zeroed after 5-window rolling validation showed negative IC across all windows. Remaining weight allocated to RSRS/Mom/Flow/RSI_Mom with ICIR²-optimized ratios. RSRS MA20 trend filter dampens z_rsrs by 50% when MA20 is declining. Portfolio stickiness penalizes new entries to reduce turnover.

Dynamic Weight Adjustment: The financial quality weight floats between 0.10 and 0.40 based on market timing signals (when the market is weak, quality drops to 0.10, releasing more weight to technical factors; during oversold bounces, quality rises to 0.40 for enhanced defense).

Weak Market Reversal Mode: When the market timing signal is < -0.3 (weak/overheated), the momentum (Mom) factor automatically switches to a short-period reversal signal (negative of 5-day returns), enabling the system to capture mean-reversion opportunities when trends fail. In weak markets, the four-quadrant definitions change to: Q1 = capital inflow + oversold price, Q2 = capital inflow + price rebound, Q3 = capital outflow + overbought price, Q4 = capital outflow + price decline.

RSRS = regression window days, Flow = share lookback days, Mom = momentum lookback days, Quality = financial quality factor, Efficiency = intraday efficiency factor, RSI_Mom = RSI momentum factor, H = optimal prediction period.

☑ Model Generalization & Backtest Results (V8)

The optimized preset was validated on an expanded 32-ETF pool covering 15 industry sectors, 8 cross-border (QDII), and 7 bond ETFs. Rolling 3-month validation with 8 prediction windows confirmed robust out-of-sample performance:

Factor Attribution (last 5 trading days, top-5 vs bottom-5 return spread):

• RSRS: Cumulative +63% — dominant alpha contributor, measuring support/resistance structural stability
• Momentum: Cumulative +22% — secondary contributor, capturing trend persistence
• Flow: Cumulative -15% — sometimes negative during volatile regimes (W5 breakdown: ICIR=-1.18)
• RSI_Mom: Cumulative -25% — negative in recent window, needs monthly review

W5 Regime Break: The Feb-Mar 2026 window showed negative ICIR (-0.71) across RSRS/Flow/Mom, but portfolio excess remained +0.70% (within -10% safety floor). The Flow factor was the primary drag (ICIR=-1.18). This confirms the model's robustness: even when factor ranking fails, the portfolio construction still limits downside.

Recommendation: Monthly factor effectiveness review is essential. If any factor shows ICIR < 0 for two consecutive months, consider reducing its weight or temporarily disabling it.

⚠ Disclaimers

• This analysis is for educational and research purposes only and does not constitute investment advice.
• Factor effectiveness fluctuates with changes in market style; historical IC does not guarantee future performance.
• IC analysis uses T+1 delay to eliminate look-ahead bias, but actual trading must account for slippage and impact costs.
• The sector ETF pool covers 32 instruments; cross-sectional IC statistics are more robust than the original 17-ETF pool, but conclusions should still be treated with caution.
• Capital flow data is based on ETF share changes and may contain noise from creation/redemption arbitrage.
• The financial quality factor is calculated from representative industry constituent stocks, not actual ETF holdings, so tracking error exists. Constituent stocks can be updated after each quarterly report.
• Financial data is inherently lagging (quarterly report disclosure delays 1-2 months); F_Quality is a slow-moving variable suitable for medium-term allocation reference.
• It is recommended to combine fundamental analysis and market environment for comprehensive judgment.