Shu and Mulvey 2024 Dynamic Factor Allocation
A 2024 Princeton ORFE working paper (arXiv:2410.14841) by Yizhan Shu and John M. Mulvey that applies regime-switching signals to dynamic factor allocation. It is one of the most carefully constructed pieces of pro-regime evidence in this vault, which is why it is treated here as a steelman candidate. The asset universe is deliberately realistic: seven long-only US equity indices — the market plus six style factors (value, size, momentum, quality, low volatility, growth) — each tracked by a heavily traded smart-beta ETF (VLUE, SIZE, MTUM, QUAL, USMV, IWF and a market index ETF) with assets under management in the billions and expense ratios near 15bp. The study explicitly rejects the academic long-short factor portfolios that are “far from practical investment opportunities” in favour of instruments an investor can actually buy.
The method runs in two stages. First, a sparse statistical jump model (SJM) identifies two regimes — bull and bear active performance — for each factor, using roughly twenty features built from factor active returns (EWMA returns, RSI, stochastic oscillator, MACD over three windows; downside deviation; active beta) plus market-environment variables (market return, VIX, 2-year yield, 10y-2y slope). Crucially the SJM is a regime classifier, not a forecaster: the paper repeatedly stresses, citing Nystrup et al., that the goal is “not to predict regime shifts … but to identify when a regime shift has occurred, and then benefit from the persistence.” Second, the per-factor regime inferences become relative views in a Black-Litterman model, producing a long-only, fully-invested allocation across the seven indices, benchmarked to an equal-weight (1/N) portfolio rebalanced quarterly.
The headline result is the steelman’s strongest single number. Over an out-of-sample test from 2007 to 2024, with a 5bp transaction cost charged on both buys and sells and a one-day delay between regime inference and rebalancing, the dynamic allocation lifts the information ratio from 0.05 (the equal-weight benchmark’s IR versus the market) to roughly 0.44, and achieves an IR of about 0.4-0.5 measured directly against the EW benchmark. Maximum drawdown relative to the market falls from -10.3% to as low as -5.9%, and the absolute Sharpe ratio improves roughly monotonically as the tracking-error budget is raised from 1% to 4%, at a one-way turnover near 500% that the authors call “within an acceptable range for active management.” Each of the six single-factor long-short evaluation strategies also earns a positive Sharpe (0.16 to 0.39). This clears genuine out-of-sample testing, realistic costs, a sensible benchmark and drawdown metrics on an investable universe — the profile of a credible result.
Three honest caveats keep it at moderate, not strong. First, the edge is small in absolute terms: active return over the equal-weight benchmark is only about 0.4-1.9% per annum depending on the tracking-error target, and the paper itself notes the equal-weight factor portfolio’s own outperformance “nearly disappeared post-2022” — the result is an information-ratio improvement on a thin active-return base, not a large standalone alpha. Second, the jump penalty is tuned to the evaluation objective: the SJM hyperparameters are chosen by time-series cross-validation to maximise the Sharpe of the hypothetical single-factor long-short strategy, the same kind of design-choice circularity flagged for Shu Yu and Mulvey 2024, so the margin should not be over-read as clean unconditioned alpha. Third, and decisively for the vault’s grading rubric, there is no independent replication: this paper sits inside the same Kolm-Lindström-Mulvey-Nystrup jump-model research network as Shu Yu and Mulvey 2024, Aydınhan Kolm Mulvey Shu 2024, Cortese Kolm Lindström 2023 and Bosancic, Nie & Mulvey 2024, and no fully independent group has reproduced a costed, out-of-sample regime-allocation result of this kind. It is also a backtest, so the Backtest-to-Live Performance Gap applies.
Net assessment: this is the cleanest pro-regime study the vault has found and it is real evidence that regime classification can add measurable value to factor allocation after costs — but it is a moderate result (profitable, costed, validated, but limited robustness and no replication), not a strong one, and it confirms rather than overturns the vault’s central finding that regime models earn their keep through risk control and allocation discipline, not large directional alpha.
Shu and Mulvey 2024 Dynamic Factor Allocation [reports_profitability] Statistical Jump Model Shu and Mulvey 2024 Dynamic Factor Allocation [tests_strategy] Style Factor Rotation Shu and Mulvey 2024 Dynamic Factor Allocation [includes_costs] Transaction Costs and Slippage Shu and Mulvey 2024 Dynamic Factor Allocation [supports] Steelman — Best Case for Markov Trading Models Shu and Mulvey 2024 Dynamic Factor Allocation [relates] Shu Yu and Mulvey 2024
Connections
- Statistical Jump Model — proposes_model, uses the sparse jump model as the regime classifier, source: https://arxiv.org/html/2410.14841v1
- Style Factor Rotation — tests_strategy, dynamic allocation across six equity style factors, source: https://arxiv.org/html/2410.14841v1
- Steelman — Best Case for Markov Trading Models — reports_profitability, the cleanest costed OOS pro-regime result, source: https://arxiv.org/abs/2410.14841
- Shu Yu and Mulvey 2024 — relates, sibling paper from the same Princeton ORFE jump-model strand, source: https://arxiv.org/abs/2402.05272
- Shu Yu and Mulvey 2024 Dynamic Allocation — relates, companion paper on asset-specific regime forecasts, source: https://arxiv.org/html/2406.09578v1
- Tactical Asset Allocation — detects_regime, regime signals feed a Black-Litterman tactical allocation, source: https://arxiv.org/html/2410.14841v1
- Transaction Costs and Slippage — includes_costs, 5bp two-way cost plus one-day inference delay, source: https://arxiv.org/html/2410.14841v1
- Out-of-Sample Backtesting — compares_benchmark, expanding-window OOS test 2007-2024 vs equal-weight benchmark, source: https://arxiv.org/html/2410.14841v1
- Maximum Drawdown — reports_profitability, drawdown vs market cut from -10.3% to -5.9%, source: https://arxiv.org/html/2410.14841v1
- Backtest-to-Live Performance Gap — lacks_live_evidence, result is a backtest with no live confirmation, source: https://arxiv.org/abs/2410.14841
- Bosancic Nie Mulvey 2024 — relates, prior regime-aware factor-allocation paper from the same group, source: https://arxiv.org/html/2410.14841v1
- US Equity Market — trades_market, source: https://arxiv.org/html/2410.14841v1