Strategies

SuperMind 周线均线金叉与日线 KDJ 共振：六只美股行业 ETF 固定规则实验

Outcome: Abandoned

SectorEtfWeeklyCrossDailyKdj

Outcome Summary

This strategy turned a SuperMind weekly-MA golden cross plus daily KDJ recipe into fixed rules and tested them on six sector ETFs over 24.5 years of daily data. It traded 193 times with low risk (5.66% max drawdown, beta 0.009) and its average trade return cleared the USEQ cost floor, but it earned only +2.32% in total, with Sharpe 0.099 and a signal IC that could not be told apart from noise. The analyst found that 2020 alone produced more than all of the profit and that only 44% of years were positive, and abandoned it after one iteration as no_edge. Risk review, promotion and paper trading were never reached.

Strategy report

Hypothesis

在六只固定行业 ETF 的 1-DAY 数据上，检验周线趋势刚转强时，日线振幅与 KDJ 金叉是否能识别随后数日至数周的上涨。原文代码没有落实振幅筛选、周线均线交叉和刚形成交叉，而且 KDJ 平滑与公式不一致；本实验采用文字机制与递归平滑公式，明确补足振幅定义、交叉有效期和退出规则。无需历史选股名单、基本面或补充数据。所有收益预期均为待检验假设。

保留来源的股票趋势与摆动指标共振机制，避免强行移植成已饱和的 BTC 期货策略。固定行业 ETF 篮子降低个股选择与历史成员数据需求；与失败的行业动量排名多空策略不同，这里不排名、不做空、不配置债券。corpus_concentration 为空，无法证明这些标的低于 15% 集中度。该实验改善多标的覆盖，但仍增加已过多的单场所、只做多策略；这一偏离是保持用户来源机制的明确代价，不能称为新的结构性套利。

premise_falsified (code: no_edge). I am not abandoning because significance is insufficient; the 'inconclusive' outcome on its own would not justify abandoning. I am abandoning because a 24.5-year, 193-trade sample shows the stated mechanism is measurably absent. The weekly SMA(5/10) golden cross plus fresh daily KDJ cross plus amplitude ≥1% on six sector ETFs (XLI/XLB/XLE/XLF/XLV/XLY) earned +2.32% in total (CAGR about 0.09%, Sharpe 0.10, PF 1.12, IR -0.32). 2020 alone carried 192% of that profit; excluding 2020 the rule loses money over the other 24 years. Only 11 of 25 years are positive (44%), which fails the 60% calendar-consistency validity floor (in soak, treated as hard). That floor cannot be waived, so a low-confidence paper slot is not available. Fee-floor comparison: avg_trade_return_pct is 0.125% against the 0.05% USEQ floor, so fees did not cause this; the gross signal has no information (forecast IC 0.033, t=0.57). The frozen plan has no tunable parameters, so neither optimization nor an exit-only iteration can change this. It matches the prior sector-ETF failures (Post-Lasso sector rotation: PF 1.000, no_edge; skipped-month sector momentum: no_edge). The verdict covers this fixed-rule daily/weekly sector-ETF translation of the SuperMind recipe, not KDJ or MA crosses on other universes.

Implementation

Fixed long-only daily experiment on six US sector ETFs. Entry requires a completed-week SMA5/SMA10 golden cross recognized within the latest ten sessions, a fresh daily K-over-D golden cross, daily high-low amplitude of at least 1% of the previous close, and K below 80. Recursive KDJ uses a nine-session range and smoothing periods of three. Exit on a daily bearish KDJ cross, K reaching 80, a nonpositive completed-week SMA spread, twenty held sessions, or a venue-managed 5% stop. Whole-share positions risk at most 0.5% of equity and deploy at most 10% per ETF, with a 60% gross basket cap.

Implements the declared instruments and frozen parameter plan without ranking, shorts, leverage amplification, supplementary feeds, or additional filters. Completed weekly closes are incorporated on the first session of the following ISO week, conservatively avoiding partial-week information; the ten-session cross window includes that recognition session. Daily amplitude is explicitly defined as 100*(high-low)/previous_close. K and D are recursively smoothed from 50, with J=3*K-2*D; K-D remains the continuous signal. All six trade decisions use exactly aligned daily close timestamps, and no future extra bar contributes. Position and regime bookkeeping uses scalar attributes per ETF; bounded indicator buffers rebuild from warmup. Protective stops are standalone because the hypothesis specifies no fixed take-profit. The conjunction may yield fewer trades than the preregistered target; its fixed rules are preserved for the pipeline to evaluate. The required outbox file could not be written because this session has read-only filesystem access; the complete task-bound JSON is returned here.

Verification Results

The sandbox produced 53 closed positions across 1,534 trading sessions. The fixed conjunction may leave limited samples for subsequent validation; assess this using full history without changing the preregistered rules.

Backtest Review

Sharpe: 0.10
Total return: 2.32%
Max drawdown: 5.66%
Trades: 193
Win rate: 48.7%
Profit factor: 1.12

The trades match the mechanism. All 193 closed positions are long (0 shorts), holding averages about 5.4 sessions, and sizing is the stated fixed 10% of equity per leg (min of 0.10, 0.5%/5% risk, 0.6/6). The code implements the preregistered completed-week SMA(5/10) cross with a 10-session validity window, a fresh daily KDJ golden cross, amplitude >= 1% and K < 80 on the six sector ETFs.

The sample is reliable and long: 1999-06 to 2024-01, about 24.5 years, all six legs aligned with zero missing bars, accounting reconciled, 0% end-unrealized.

Numerically viable for USEQ: avg_trade_return_pct 0.1245% against the 0.05% floor, above the preferred 0.10%. PF 1.116 and Sharpe is positive. Max drawdown is 5.66% against the pre-registered 15%.

Market beta is not the explanation. Beta is 0.009, exposure 11.9% and drawdown 5.7%, while the equal-weight sector basket fell about 50% in 2008. This is not a lesson-170 beta book.

The edge cannot be told apart from zero. Sharpe is 0.099 with a 95% CI of [-0.28, 0.51] and PSR 0.56. The signal-IC diagnostic gives 0.033 (t=0.57, rejected_no_information). My rough per-trade t-statistic is about 0.7: mean 0.12% against a per-trade standard deviation of about 2.5% over 193 trades.

The economics are negligible: +2.3% total over 24.5 years, Calmar 0.017, information ratio -0.32 against the equal-weight basket, and a maximum drawdown duration of 6857 days. Only 12 of 25 calendar years were positive, and 2020 (+4.68%) carries more than the entire net profit. The yearly-consistency validity floor is at risk.

Market impact takes 31% of gross PnL even at 10% position sizes.

QA's concern is partly confirmed. The full history gives 193 trades (about 8 per year, not the 53 seen in the sandbox), but the last-20% holdout would get only about 40 trades. A positive-net-edge HAC test at this effect size will very likely come back inconclusive.

The plan is frozen (tunable is empty), so 'optimization' here means a frozen walk-forward/holdout evaluation, not a parameter search. Nothing will be fitted.

Analysis

Implementation matches the preregistered rules: long-only, completed-week SMA(5/10) golden cross plus fresh daily KDJ golden cross, amplitude filter, KDJ/weekly/time exits. No mechanism mismatch.

Per-trade return of 0.125% clears the USEQ cost floor (0.05%), so fees are not the problem. Commission is zero; impact is 31% of gross.

Risk is small: max drawdown 5.66% against the preregistered 15%, about 10% of capital per position, 12% exposure. Beta 0.009, so this is not a closet long.

Economically there is almost nothing here. Total return is +2.32% over 24.5 years (CAGR about 0.09%), Sharpe 0.099 (95% CI -0.28 to 0.51), PF 1.116, Calmar 0.02. The drawdown took 6857 days (essentially the whole sample) to recover. Information ratio is -0.32 against the equal-weight sector basket.

All of the profit comes from one year. 2020 contributed +4.68%, which is 192% of the total. Excluding 2020, the rule loses about 2.4% over 24 years. Only 11 of 25 calendar years are positive (44%). That is below the 60% validity floor, which is in its soak period but should be treated as hard.

The frozen-plan walk-forward reuses the same parameters in every window (no tunable parameters), so the outer windows are a chronological replay. Their OOS Sharpes are scattered and mostly at or below zero (e.g. -2.83, -3.76, -2.01, -2.14, -2.84, -2.94, -2.91), with 0-5 trades per window. Inner IS Sharpe turns negative after 2012.

The executed-forecast IC is 0.033 (t=0.57), which the platform's signal diagnostic rejects as 'rejected_no_information'.

Preregistered predictions: 1 of 4 met. 193 trades vs 200; return 2.3% vs 12%; avg trade 0.12% vs 0.30%. Only drawdown (5.7% vs 15%) was met.

not beta-driven; simply no active edge

beta 0.009, IR -0.32 vs equal-weight basket

validity floor in soak, treated as hard

advisory only

USEQ floor; 2x economics floor 0.10 also passes

Outcome Summary

Fixed MA-cross plus KDJ timing rules on liquid US sector ETFs show no measurable gross edge even over a long, clean sample. This adds to earlier sector-ETF no_edge results, so future sector-ETF hypotheses need a different information source, not another variant of these technical triggers.

The backtest review sent it on to a frozen walk-forward evaluation; the plan had no tunable parameters. The analyst then abandoned it as premise_falsified (no_edge). 2020 produced 192% of the total profit, the rule lost money in the other 24 years combined, and only 11 of 25 years were positive (44%), which fails the 60% calendar-consistency floor. That floor cannot be waived, so a low-confidence paper slot was not available.

Long-only entries on six US sector ETFs (XLI, XLB, XLE, XLF, XLV, XLY) on daily bars, taken when a completed-week SMA(5/10) golden cross from the last 10 sessions lines up with a fresh daily KDJ golden cross, daily amplitude of at least 1% and K below 80. This was a fixed-rule version of a SuperMind recipe.

Over 1999-06 to 2024-01 (about 24.5 years) it made 193 long trades. Total return was +2.32% (CAGR about 0.09%) with Sharpe 0.099 (95% CI -0.28 to 0.51), profit factor 1.116, average trade return 0.125% of notional, max drawdown 5.66% and information ratio -0.32 against the equal-weight sector basket. The signal IC was 0.033 (t=0.57), which the platform's signal diagnostic rejected as carrying no information.

Analysis

(supermind量化-)振幅大于1、周线MA5金叉MA10、KDJ刚形成金叉_

Backtest and paper results are hypothetical. Trading involves risk of loss.