Alpha 掘金系列之二十：热门概念板块 AI 预测与概念龙头识别

Quantitative Models and Construction Methods 1. Model Name: TimeMixer-based Hot Concept Index Rotation Strategy - Model Construction Idea: Aggregate individual stock Alpha factors into concept indices and construct a rotation strategy based on the aggregated factor scores[4][39][44] - Model Construction Process: 1. Develop Alpha factors at the individual stock level[39] 2. Aggregate these Alpha factors into concept indices using equal weighting, as Wind Hot Concept Indices are equally weighted[4][44] 3. Perform factor IC tests on the aggregated indices, achieving an IC mean of 7.27%[44][49] 4. Construct a weekly rotation strategy by selecting the top 10 concept indices with the highest model scores and allocating them equally[4][51] 5. Backtest the strategy from January 4, 2019, to August 29, 2025, with a transaction cost of 0.1% per side[4][51] - Model Evaluation: The strategy demonstrates strong performance with consistent positive excess returns over the benchmark in all years[4][51][57] 2. Model Name: Alpha Stock Portfolio Based on Hot Concept Index Rotation Effect - Model Construction Idea: Enhance the operability of the rotation strategy by selecting a smaller number of representative stocks from the concept index components[64] - Model Construction Process: 1. Use weekly signals from the concept index rotation strategy[64] 2. Rank the component stocks of the indices based on the TimeMixer-enhanced machine learning Alpha factor[64] 3. Select the top 20 stocks and construct an equally weighted portfolio[64] 4. Backtest the strategy from January 2, 2019, to August 29, 2025, with a transaction cost of 0.1% per side[64] - Model Evaluation: The strategy achieves stable positive excess returns in most years but shows no significant advantage over traditional machine learning methods[64][65] 3. Model Name: Leading Stock Portfolio Based on Hot Concept Index Rotation Effect - Model Construction Idea: Identify core leading stocks within concept indices using the Free Cash Flow to Enterprise Value (FCF2EV) factor[70] - Model Construction Process: 1. Define the FCF2EV factor as the ratio of free cash flow (FCF) to enterprise value (EV), where FCF = (1-t) * EBIT + depreciation/amortization - CapEx - net working capital changes, and EV = market capitalization + total debt - cash[70] 2. Select the top 2 stocks with the highest FCF2EV values from each concept index's components weekly[70] 3. Construct an equally weighted portfolio and backtest the strategy from January 2, 2019, to August 29, 2025, with a transaction cost of 0.1% per side[70] - Model Evaluation: The strategy performs well, achieving positive excess returns in all years, though it underperforms in growth-dominated years due to the value and large-cap bias of the FCF2EV factor[70][75] --- Model Backtesting Results 1. TimeMixer-based Hot Concept Index Rotation Strategy - Annualized Excess Return (vs. CSI All Share Index): 18.06%[51][52] - Annualized Excess Return (vs. Equal-weighted Hot Concept Index): 9.02%[51][52] - Information Ratio (IR, vs. CSI All Share Index): 1.73[51][52] - Information Ratio (IR, vs. Equal-weighted Hot Concept Index): 0.76[51][52] - Maximum Excess Drawdown (vs. CSI All Share Index): 9.97%[51][52] - Maximum Excess Drawdown (vs. Equal-weighted Hot Concept Index): 21.74%[51][52] 2. Alpha Stock Portfolio Based on Hot Concept Index Rotation Effect - Annualized Excess Return: 11.34%[64][65] - Information Ratio (IR): 0.79[64][65] - Maximum Excess Drawdown: 22.87%[64][65] 3. Leading Stock Portfolio Based on Hot Concept Index Rotation Effect - Annualized Excess Return (vs. CSI All Share Index): 20.63%[71][73] - Annualized Excess Return (vs. Equal-weighted Hot Concept Index): 11.52%[71][73] - Information Ratio (IR, vs. CSI All Share Index): 1.61[71][73] - Information Ratio (IR, vs. Equal-weighted Hot Concept Index): 0.88[71][73] - Maximum Excess Drawdown (vs. CSI All Share Index): 21.65%[71][73] - Maximum Excess Drawdown (vs. Equal-weighted Hot Concept Index): 21.02%[71][73] --- Quantitative Factors and Construction Methods 1. Factor Name: Momentum Factor (Monthly and Weekly) - Factor Construction Idea: Evaluate the momentum characteristics of the Hot Concept Indices[23] - Factor Construction Process: 1. Calculate the Rank IC as: $ RankIC_t = corr(Rank(X_{t,m}), Rank(r_{t+1,m})) $ where $X_{t,m}$ is the factor value, and $r_{t+1,m}$ is the next period's asset return[23] 2. Divide assets into 10 groups based on factor values and construct long-short portfolios (L-S portfolios) by going long on the top group and short on the bottom group[23] - Factor Evaluation: The monthly momentum factor shows weaker monotonicity in long-short portfolio returns compared to the weekly momentum factor[23][26] 2. Factor Name: Free Cash Flow to Enterprise Value (FCF2EV) - Factor Construction Idea: Identify leading stocks with strong financial health and risk resistance within concept indices[70] - Factor Construction Process: 1. Define FCF2EV as: $ FCF2EV = \frac{FCF}{EV} $ where $ FCF = (1-t) * EBIT + Depreciation/Amortization - CapEx - Net Working Capital Changes $ and $ EV = Market Capitalization + Total Debt - Cash $[70] 2. Calculate FCF2EV for each stock in the concept index components and rank them[70] - Factor Evaluation: The FCF2EV factor aligns with value and large-cap styles, performing well in most years but underperforming in growth-dominated years[70][75] --- Factor Backtesting Results 1. Momentum Factor - Monthly Momentum Factor: - IC Mean: 1.35%[26] - Long-only Annualized Excess Return: 22.17%[26] - Long-only IR: 0.36[26] - Long-short Annualized Excess Return: 11.13%[26] - Long-short IR: -0.31[26] - Weekly Momentum Factor: - IC Mean: 2.38%[26] - Long-only Annualized Excess Return: 24.03%[26] - Long-only IR: 0.53[26] - Long-short Annualized Excess Return: 12.21%[26] - Long-short IR: -0.26[26] 2. Free Cash Flow to Enterprise Value (FCF2EV) Factor - Annualized Excess Return (vs. CSI All Share Index): 20.63%[71][73] - Annualized Excess Return (vs. Equal-weighted Hot Concept Index): 11.52%[71][73] - Information Ratio (IR, vs. CSI All Share Index): 1.61[71][73] - Information Ratio (IR, vs. Equal-weighted Hot Concept Index): 0.88[71][73] - Maximum Excess Drawdown (vs. CSI All Share Index): 21.65%[71][73] - Maximum Excess Drawdown (vs. Equal-weighted Hot Concept Index): 21.02%[71][73]