Strategies

Size-Style Relative-Strength Trend Rotation, Dollar-Neutral SPY vs IWM (USEQ 1-DAY) — Port of CICC FOF #8 Large/Small Style Rotation

Outcome: Abandoned

SizeStyleRsTrendRotationSpyIwmUseq

Outcome Summary

After one iteration, the strategy tested daily SPY/IWM relative-strength rotation using approximately equal opposing leg notionals. Input audits passed across 5,940 aligned sessions, but the baseline lost 14.07%, with Sharpe 0.027 and profit factor 0.945. Backtest review abandoned it for negative expectancy before optimization or later reviews. The result concerns the price-only adaptation, leaving CICC's unavailable fund-allocation lead signal untested.

Strategy report

Hypothesis

Classification: (a) a concrete idea from an external source. Source: CICC 'FOF研究系列之八：基金择时、选股和风格轮动' (2018-08-21, via BigQuant wiki), section '基于基金配置动量效应的风格轮动策略'. The article trades size-style momentum. When the large-vs-small relative-strength (RS) indicator trends up, it favours the large-cap index (CSI 300). When RS trends down, it favours the small-cap index (CSI 1000). A long-short version holds the favoured style against the other, with an average hold of about one month. The article's first strategy (an FOF built from fund managers with stock-picking skill) cannot be tested here because we have no mutual-fund NAV or holdings data, so it is dropped. Two adaptations, stated explicitly. (1) Market: CNEQ holds only about 100 CSI 300 constituents. It is long-only and has no CSI 1000 or other small-cap index instrument, so the A-share pair cannot be built. The same large-vs-small rotation is tested on USEQ with SPY.USEQ (large cap) against IWM.USEQ (small cap). Both have decades of daily history, can be shorted on Reg-T margin and trade commission-free. (2) Signal: the article's leading input is the aggregate change in mutual-fund style allocation. It reports a 0.75 correlation with the style-index relative return (0.81 at a 60-day lead). That feed does not exist in our data, so the strategy uses the price-only RS trend that the fund signal tracks: RS = close(SPY)/close(IWM). This tests the style-momentum premise, not the fund-positioning signal, and any verdict should say so. Construction: dollar-neutral long/short, with each leg at 50% of equity (gross 1.0x). Decisions are taken on the daily close and fill at that close. Next-session-open fills are unsupported on USEQ, so the strategy does not claim the overnight gap. Dividends are not in PnL. Long+short, multi-instrument, session venue, no supplementary data.

USEQ is 6.6% of the corpus and long-short only 18.5%. This is a style-factor mechanism on non-BTC instruments, not another crypto momentum variant. CNEQ cannot host the CSI300-vs-CSI1000 pair (no small-cap instrument, long-only). USEQ keeps the exact large-vs-small mechanism with shortable ETFs and long history. There are only three tunables because overfitting has been a recurring failure.

negative_expectancy: this default price-only SPY/IWM style-momentum implementation produces negative net returns and PF below 1 on a functioning, multi-decade backtest, so it does not merit optimization expenditure. This result does not reject CICC's unavailable fund-allocation lead signal or establish that every alternative parameter configuration lacks an edge. Single-leg buy-and-hold comparison is inappropriate because benchmark_meaningful is false.

Implementation

Dollar-neutral size-style rotation on USEQ daily bars. RS = ln(SPY/IWM). The signal is the change in the ma_len-session moving average of RS over slope_window sessions. Above +deadband: long SPY / short IWM. Below -deadband: short SPY / long IWM. Each leg is 50% of equity in whole shares (gross 1.0x). There is a pair stop at an 8% adverse RS move. After a 10-session minimum hold, the pair exits on an opposite-side signal (the reverse entry is taken on the next session) or after 10 consecutive sessions inside the deadband. Fills are at the daily close.

This ports the CICC FOF #8 large/small style-momentum rotation to SPY vs IWM, because CNEQ has no small-cap instrument and cannot short. It uses the price-only RS trend, so it tests style momentum and not the mutual-fund allocation signal. Tunable parameters are exactly the plan's ma_len, deadband and slope_window. The plan's fixed values are implemented as defined: min_hold_bars=10, pair_stop_pct=0.08 (measured as an adverse log-RS move from entry, i.e. roughly long-leg return minus short-leg return), neutral_exit_bars=10 and leg_equity_fraction=0.5. Interpretation choices: after a pair stop the same direction is blocked for min_hold_bars sessions, so the stop cannot re-enter immediately; a flip exits and re-enters on the next session. Each leg is its own order and position. Position state is kept as plain scalars. Indicator buffers are rebuilt by warmup replay. IWM is read only when its bar timestamp equals the SPY bar timestamp. Note: ma_len=120 with slope_window=60 needs 180 sessions, more than the plan's 130-bar warmup, so those combinations stay flat for their first ~50 sessions.

Verification Results

Optional: if the risk officer's drawdown floor binds, scale leg_equity_fraction or pair_stop_pct. It is fixed in the plan, so leave it as-is for the baseline.

Each leg is equity*0.5 and is recomputed from current equity at each entry. Gross exposure is capped at 1.0x by construction, under the USEQ 2x Reg-T limit, and there is no dynamic scalar. The pair stop is an 8% adverse log-RS move, which costs about 4% of equity per stop. That is above the 1-2%-of-equity per-trade risk guideline (L148). This is not a blowup risk, because the design has a flat state, a stop and bounded gross.

The analyst verdict should say it tests price-only size-style momentum, not the fund-positioning signal.

This is a disclosed adaptation, not a defect: the market moves from CSI300/CSI1000 to SPY/IWM, and the signal is the price RS trend instead of the mutual-fund style-allocation change. The hypothesis states both changes explicitly, and the code matches the adapted hypothesis.

Consider raising warmup_bars to 180 in the evaluation plan.

min_bars_required returns 2, so readiness is handled internally (ma_len + slope_window sessions). The developer correctly notes that ma_len=120 with slope_window=60 needs 180 sessions, more than the 130-bar warmup, so those optimizer draws stay flat for about 50 extra sessions after each walk-forward reset. Results are slightly understated, but nothing is incorrect.

The sandbox ran about 1,524 sessions and produced Sharpe 0.17, PF 1.12, 14.4% max drawdown and a 0.48% average return per closed position over 40 legs (about 20 pairs). The PSR is 0.65 and the Sharpe CI of [-0.60, 0.95] spans zero. Borrow cost was 1.4% of equity, more than half of the excess over flat. Only about 20 independent pair decisions were made in roughly 6 years, so the full-history backtest needs to show more occurrences before any conclusion. The test also uses only the price RS trend, not the CICC fund-allocation lead signal, so a null result falsifies price-only US size momentum, not the source signal. Spread cost on USEQ (about 0.02-0.05% round trip per leg) is small next to an average pair hold of about 94 days, so turnover economics are not the concern.

Backtest Review

Sharpe: 0.03
Total return: -14.07%
Max drawdown: 23.16%
Trades: 146
Win rate: 47.3%
Profit factor: 0.94

Input audits passed across 5,940 aligned sessions.

Spot-checked trades have simultaneous opposing SPY/IWM legs with approximately equal entry notionals, consistent with the paired construction.

The baseline lost 14.07% over roughly 23.6 years; profit factor is 0.945 and average net leg return is -0.184% of notional.

146 closed legs represent approximately 73 completed pairs, not 146 independent decisions. Average holding duration is about 96 days.

QA's weak-economics concern is borne out by the full-history losses. No fixable implementation defect or measured profitable parameter region has been established.

USEQ economic reference; rejection rests on measured negative net expectancy, not a promotion gate.

Iteration History

Each leg is equity*0.5 and is recomputed from current equity at each entry. Gross exposure is capped at 1.0x by construction, under the USEQ 2x Reg-T limit, and there is no dynamic scalar. The pair stop is an 8% adverse log-RS move, which costs about 4% of equity per stop. That is above the 1-2%-of-equity per-trade risk guideline. This is not a blowup risk, because the design has a flat state, a stop and bounded gross.

Outcome Summary

This baseline's failure applies to the tested price-only implementation and does not reject the unavailable fund-allocation signal or establish that every alternative parameter configuration lacks an edge.

Backtest review issued an abandon verdict because the functioning multi-decade backtest showed negative net expectancy, with no established fixable implementation defect or measured profitable parameter region. Optimization and subsequent analyst and risk reviews were not reached.

It tested price-only large-versus-small-cap momentum through daily dollar-neutral SPY/IWM long-short rotation, adapting CICC's style-rotation premise without its fund-allocation lead signal.

The baseline returned -14.07% over roughly 23.6 years, with Sharpe 0.027 and 23.16% maximum drawdown. It recorded 146 closed legs, approximately 73 completed pairs, with profit factor 0.945 and average net leg return -0.184% of notional.

Analysis

FOF研究系列之八：基金择时、选股和风格轮动中金公司_20180821 - BigQuant量化交易

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