SignalTrace

Statistical Jump Model

John Montgomery6 min read

Statistical Jump Model

The statistical jump model (JM) is a non-parametric regime-identification model that extends k-means clustering of temporal features by adding an explicit jump penalty charged at every change of hidden state. The penalty hyperparameter directly controls regime persistence, addressing the HMM’s tendency to flip states too often. Unlike traditional Markov-switching models it enforces persistence through this penalty rather than through a likelihood with assumed return distributions, which its proponents argue makes it more robust to the mis-estimation and mis-specification problems that afflict HMMs on non-normal, time-varying financial data. The method was introduced by Nystrup Lindström Madsen 2020 in Expert Systems with Applications as a way of “learning hidden Markov models with persistent states by penalizing jumps” — the originating methodology paper, not a finance-application paper. Its stated advantages over standard Baum-Welch/EM estimation are concrete and testable: it learns the state sequence and parameters jointly and faster, gives explicit control over the transition rate, is less sensitive to initialisation, scales better as the number of states grows, and is robust to misspecified conditional distributions.

The model family has since been extended along two axes by overlapping author groups. The sparse statistical jump model of Nystrup Kolm Lindström 2021 adds joint feature selection via a coordinate-descent algorithm with a sparsity budget, so the model can isolate the handful of features that actually distinguish regimes in a high-dimensional, noisy feature set — a capability that has historically been infeasible for standard HMMs. The continuous statistical jump model (CJM) of Aydınhan Kolm Mulvey Shu 2024 replaces each period’s hard regime label with a probability vector over all regimes and introduces a “mode loss” penalty, giving smoother transitions and better behaviour under regime imbalance and limited data. Reference implementations of all three variants are available in the open-source jumpmodels Python library, which is a genuine point in favour of reproducibility — the methodology is open and runnable.

In the trading-relevant head-to-head of Shu Yu and Mulvey 2024 — a 0/1 equity-timing strategy on the S&P 500, DAX and Nikkei 225 over 1990-2023, with 10bp one-way transaction costs and trading-delay robustness checks — the JM-guided strategy beat the HMM-guided strategy on every index: S&P 500 return 11.2% vs 8.5% (Sharpe 0.68 vs 0.54), DAX 8.6% vs 6.4%, Nikkei 4.7% vs 2.5%, with consistently lower maximum drawdown. The decisive difference was turnover: 44%/170%/72% for the JM against 141%/246%/290% for the HMM. The JM’s edge is therefore best read as turnover and drawdown control — the calmer, more persistent regimes avoid the cost-eroding churn that destroyed the HMM’s net edge — rather than as a large new alpha source. Against a buy-and-hold benchmark the JM’s absolute return advantage is modest (S&P 500) or comes almost entirely from drawdown reduction (DAX, Nikkei).

The honest replication verdict is mixed. The JM’s methodological advantages over the HMM — controllable persistence, robustness to misspecification, lower turnover, feature selection — are confirmed by more than one paper and more than one (partly overlapping) author group: Nystrup/Lindström/Madsen, Nystrup/Kolm/Lindström, Cortese/Kolm/Lindström, and Aydınhan/Kolm/Mulvey/Shu. Independent-ish application to a new asset class exists: Cortese Kolm Lindström 2023 apply the sparse JM to cryptocurrency and confirm it produces interpretable, persistent bull/neutral/bear regimes. But every one of these papers shares at least one author with the others (Kolm and Lindström are the connective tissue; Mulvey/Shu link the Princeton strand). There is, as of 2024, no fully independent research group that has reproduced a costed, out-of-sample JM-beats-HMM trading result. Furthermore the only paper that actually grades a trading strategy — Shu, Yu & Mulvey — tuned the jump penalty by cross-validation to maximise the very 0/1 strategy being evaluated, so the JM-vs-HMM margin should not be over-read as clean out-of-sample alpha.

Net assessment for this vault: the JM is a real and useful improvement on the HMM as a regime classifier and risk filter — calmer regimes, far lower turnover, demonstrably lower drawdown — and that improvement is confirmed across the model-methodology literature. Its claimed trading edge over the HMM is plausible and internally consistent but rests on a single research network and a penalty tuned to the evaluated strategy, so it earns a moderate profitability grade, not strong: it clears costs and out-of-sample testing in one careful study, but not independent replication. Like every regime model it shares the ~25-day detection-latency limitation and is a downside-risk tool, not an alpha engine.

Statistical Jump Model [contradicts] Hidden Markov Model Regime Detection Statistical Jump Model [relates] Markov Regime-Switching Model Nystrup Lindström Madsen 2020 [defines] Statistical Jump Model Nystrup Kolm Lindström 2021 [defines] Statistical Jump Model Aydınhan Kolm Mulvey Shu 2024 [defines] Statistical Jump Model Peter Nystrup [proposes_model] Statistical Jump Model Petter Kolm [proposes_model] Statistical Jump Model Shu Yu and Mulvey 2024 [tests_strategy] Regime-Based Asset Allocation Cortese Kolm Lindström 2023 [detects_regime] Cryptocurrency Market

Connections

Sources

Graph View

Backlinks