Adaptive CUSUM Chart Overview

• Adaptive CUSUM charts are sequential change-detection schemes that adjust control limits and parameter estimates based on real-time data instead of relying on fixed thresholds.
• They incorporate methods like observation-adjusted boundaries, online estimation, self-starting techniques, and non-restarting processes to improve detection of regime changes.
• These adaptive strategies enhance performance by reducing detection delays in diverse scenarios, including heavy-tailed distributions and situations with unknown baseline parameters.

An adaptive CUSUM chart is a sequential change-detection scheme based on the cumulative sum principle in which at least one component of the monitoring rule is updated from the observed process rather than kept fixed throughout operation. In the recent literature, this adaptivity takes several distinct forms: observation-adjusted control limits, online estimation of out-of-control parameters, self-starting standardization when baseline parameters are unknown, adaptive weighting for unknown change locations, and non-restarting bounded signal processes for recurrent regime changes [(Tang et al., 2023); (Parakulum et al., 30 Aug 2025); (Bourazas, 2024); (Gandy et al., 2012)]. The common departure from the classical formulation is that the chart no longer relies solely on a fixed threshold and a fully specified post-change model.

1. Classical baseline and the meaning of adaptivity

The classical upper-sided likelihood-ratio CUSUM is written as

$$T_c(c)=\min\left\{n\ge 0:\max_{1\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\right\}, \qquad Z_i=\log\frac{p_{v_1}(X_i)}{p_{v_0}(X_i)}.$$

In the Gaussian known-parameter setting, a standard recursion is

$$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$

with $U_n$ standardized and the run length defined as the first threshold crossing (Tang et al., 2023, Lombard et al., 2019).

Adaptive CUSUM constructions modify this fixed design in different ways. Some papers replace the constant control limit $c$ by an observation-dependent boundary. Others retain the CUSUM recursion but replace unknown out-of-control parameters by estimates based on recent post-reset data. Self-starting variants update in-control mean and variance sequentially so that monitoring can begin immediately, while nonparametric variants replace model-based likelihood ratios by rank, category, or transform-based surrogates (Tang et al., 2023, Parakulum et al., 30 Aug 2025, Li, 2017, Bourazas, 2024).

This diversity of constructions implies that “adaptive CUSUM” is not a single chart but a family of design principles. A plausible implication is that adaptivity should be understood operationally: the chart changes some aspect of its own decision mechanism in response to the incoming stream rather than only accumulating fixed-form evidence.

2. Observation-adjusted thresholds and dynamic stopping boundaries

A direct form of adaptivity is to make the control limit itself depend on the observed sequence. In the CUSUM test with observation-adjusted control limits, or CUSUM-OAL, the stopping time is

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$

where $g(\cdot)$ is a decreasing function. If $g\equiv 1$, the procedure reduces to the classical CUSUM (Tang et al., 2023).

The rationale is explicit: if the observations suggest an increasing mean, the alarm threshold should be lowered. In this formulation, the adaptive chart forms a bridge between fixed-threshold CUSUM and faster likelihood-based procedures. The paper establishes the limits

$$\lim_{u\to\infty}T_c(cg_{u,r}) = T^*(r), \qquad \lim_{u\to\infty}T_c(cg_u)=T_c(c)\wedge T^*(0),$$

so the observation-adjusted family connects the classical CUSUM to modified sum-log-likelihood-ratio tests (Tang et al., 2023).

The same idea has been extended to extremely heavy-tailed sequences, again under the name CUSUM-OAL. There the adaptive boundary uses a local average

$$\hat{Z}_n(ac)=\frac{1}{j}\sum_{i=n-j+1}^n Z_i, \qquad j=\min\{n,[ac+1]\},$$

and the chart is developed for shifts in the tail parameter $\alpha$ of Pareto-type laws

$$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$0

The heavy-tail setting is essential because for $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$1 the mean and variance are infinite, so mean- or variance-based monitoring is not suitable (Tang et al., 2024).

Both versions analyze average run lengths asymptotically. For the observation-adjusted schemes, the out-of-control behavior depends on the sign of $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$2: exponential-type for small changes, square-order for medium changes, and linear-order for large changes (Tang et al., 2023, Tang et al., 2024). This identifies threshold adaptation as more than a heuristic device; it changes the asymptotic run-length regime.

3. Online estimation of the out-of-control regime

Another major adaptive strategy replaces the unknown post-change model by recursively updated estimates. In the Gaussian joint mean-and-variance problem, the ideal standardized CUSUM increment is written as

$$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$3

which is optimal when $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$4 are known. The adaptive version replaces these by estimates derived from the most recent post-reset observations, with the estimated change-point defined as the most recent time the statistic reset to zero. The method maintains eight directional CUSUMs, one for each nontrivial combination of change directions in mean and variance, and then standardizes their in-control distributions before aggregation by a maximum rule. It also provides built-in post-signal diagnostics, because the chart that crosses first indicates the likely type of change (Parakulum et al., 30 Aug 2025).

A related construction appears for real-time monitoring of bivariate time-between-events data when the two components arrive asynchronously. The incoming ordered observations are transformed by conditional probability integral transforms, then mapped to independent exponential variables

$$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$5

so that under in-control conditions they are i.i.d. $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$6. The base likelihood-ratio recursion becomes

$$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$7

but the rates $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$8 are unknown and are therefore estimated online from the current excursion since the last reset by Gamma-prior shrinkage. Eight adaptive CUSUMs are maintained for the sign combinations $$D_0 = 0, \qquad D_n = \max\{0, D_{n-1} + U_n - \delta\}, \quad n \ge 1,$$9, after which the individual statistics are put on a common null scale and combined by a maximum (Parakulum et al., 30 Aug 2025).

The same estimation logic also underlies adaptive nonparametric charts. In the distribution-free categorical framework, continuous observations are converted into left-to-right and center-outward categories, and the unknown out-of-control multinomial probabilities are estimated online from the post-reset segment through

$U_n$0

The resulting four directional charts $U_n$1 are then combined by a maximum, again with built-in post-signal diagnosis (Li, 2017).

In continuous time, adaptivity can be expressed as sequential estimation of a hazard increase. The CGR-CUSUM for survival outcomes uses

$U_n$2

and constructs

$U_n$3

The hazard increase is therefore estimated from the data rather than specified in advance (Gomon et al., 2022).

Across these examples, the adaptive step is not the cumulative sum itself but the replacement of a fixed out-of-control specification by a recursively updated surrogate. This suggests that much of the modern literature treats adaptive CUSUM as a plug-in sequential likelihood-ratio architecture.

4. Self-starting, weighted, Bayesian, and rank-based formulations

Self-starting CUSUMs address the case in which in-control parameters are unknown and Phase I calibration is unavailable or undesirable. Hawkins’ self-starting scheme begins with a warmup sample of size $U_n$4, updates the mean and variance recursively, and transforms the resulting standardized residuals into independent exact $U_n$5 variables through

$U_n$6

With these transformed $U_n$7, the self-starting chart can use the same control limits as the known-parameter CUSUM, provided fresh warmup observations are regenerated after each signal (Lombard et al., 2019).

The recent comparative study of self-starting location charts places this frequentist Self-Starting CUSUM alongside the Bayesian Predictive Ratio CUSUM. In the Bayesian version, the process model is

$U_n$8

and the chart accumulates a log predictive ratio between in-control and shifted posterior predictive densities. Both schemes are self-starting in the sense that they estimate the unknown parameters sequentially from the start of monitoring, but they differ in how incremental evidence is generated: transformed frequentist residuals for SSC and predictive log ratios for PRC (Bourazas, 2024).

Bayesian adaptivity is broader than the self-starting normal model. In profile monitoring, a Bayesian CUSUM is written as

$U_n$9

with posterior and posterior predictive distributions updated as new data arrive. The selected loss function—SELF, PLF, or LLF—changes the Bayes estimator and therefore changes the chart center and effective reference structure (Mitchell et al., 2020).

Adaptive weighting is a different mechanism. In one-change problems, the $c$0-weighted CUSUM uses weights $c$1, and the adaptive version estimates the weight through

$c$2

where $c$3 is Lipschitz continuous and satisfies $c$4. Under $c$5, the adaptive statistic behaves asymptotically like the standard CUSUM; under $c$6, it approximates the oracle weight determined by the true change location (Schwaar, 2020).

Rank-based and geometric self-standardization extend adaptivity beyond parametric models. The Mann–Whitney CUSUM accumulates the standardized statistic $c$7 to detect small location shifts under a completely distribution-free framework, while the angular-data CUSUM replaces the unknown mean direction by the current estimate $c$8 and uses

$c$9

or, for concentration monitoring,

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$0

These charts update the underlying reference structure from the data stream itself rather than from a fixed parametric baseline [(Wang et al., 2013); (Lombard et al., 2017)].

5. Non-restarting adaptive signal processes and recurrent regimes

A distinct branch of the literature treats adaptivity as continuous signaling through alternating in-control and out-of-control episodes. In the non-restarting bounded CUSUM, the statistic evolves as

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$1

where $$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$2 is an upper boundary. The chart is not reset after signaling; instead, it signals continuously while above threshold. The upper boundary prevents the statistic from rising too high during a long abnormal period and helps detect a return to in-control behavior (Gandy et al., 2012).

For multiple streams, each bounded non-restarting CUSUM produces a p-value

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$3

and a pointwise false discovery rate rule such as Benjamini–Hochberg is applied at each time. The methodology is proved to control the false discovery rate under two distinct definitions of false discovery, including the adaptive definition tied to whether a stream has been in control since the last time its chart hit zero (Gandy et al., 2012).

The optimality theory for repeated switching between two known distributions uses the bounded recursion

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$4

and replaces one-shot stopping times by a signal process $$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$5. Here the lower and upper barriers make the chart a continuously updating state classifier. The main theorem shows that, for suitable thresholds and $$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$6, the resulting non-restarting signal process is optimal in a repeated-change worst-case expected-delay sense (Lau et al., 2012).

This line of work broadens the meaning of an adaptive CUSUM chart. The adaptation is not primarily parameter estimation or threshold modulation; it is the refusal to treat a signal as the end of the monitoring cycle. Instead, the chart itself becomes a persistent state process.

6. Performance criteria, diagnostics, and conceptual issues

Average run length remains the dominant performance criterion, but adaptive CUSUM papers use several variants. In-control and out-of-control ARLs are central in observation-adjusted, Bayesian, nonparametric, and joint mean/variance formulations (Tang et al., 2023, Mitchell et al., 2020, Parakulum et al., 30 Aug 2025). Short-run ratio monitoring replaces ARL by the truncated average run length,

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$7

because the production run ends after $$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$8 planned inspections (Tran, 4 Jun 2026). Other papers use average time to signal (ATS) for time-between-events data and conditional expected delay (CED) for self-starting location charts (Parakulum et al., 30 Aug 2025, Bourazas, 2024).

A major conceptual issue concerns what ARL means for self-starting charts. The conditional average run length,

$$T_c(cg)=\min\left\{n\ge 0:\max_{0\le k\le n}\sum_{i=n-k+1}^n Z_i \ge c\,g(\bar Z_n)\right\}, \qquad \bar Z_n = \frac{1}{n}\sum_{i=1}^n Z_i,$$9

varies with the warmup sample, but for a properly used self-starting CUSUM the relevant measure is the unconditional ARL

$g(\cdot)$0

because fresh warmup data are required after each restart. The literature therefore distinguishes between a mathematically legitimate conditioning argument and the performance actually experienced under correct repeated use (Lombard et al., 2019).

Another issue is the scope of the term “adaptive.” Not every optimized CUSUM design is adaptive in the online sense. The NSGA-II economic-statistical design optimizes $g(\cdot)$1, $g(\cdot)$2, and $g(\cdot)$3 for a conventional CUSUM chart by minimizing $g(\cdot)$4 and $g(\cdot)$5, but it does not dynamically adapt $g(\cdot)$6, $g(\cdot)$7, $g(\cdot)$8, or $g(\cdot)$9 during operation. It is therefore described as a multi-objective economic-statistical optimization of a fixed CUSUM chart design rather than a truly adaptive CUSUM scheme (Sandeep et al., 2024).

By contrast, short-run ratio monitoring shows that design-level adaptivity can itself be substantive when it changes the feasible operating region. The adaptive upper one-sided CUSUM-RZ$g\equiv 1$0 jointly optimizes the reference value $g\equiv 1$1 and decision interval $g\equiv 1$2 under a TARL$g\equiv 1$3 constraint through a bilevel procedure, and the inner calibration step can recover boundary cases that fixed-$g\equiv 1$4 designs cannot attain (Tran, 4 Jun 2026). This suggests a useful distinction between online adaptivity, which changes the chart while it runs, and design adaptivity, which changes how the chart is calibrated for the intended deployment horizon.

Taken together, these developments show that adaptive CUSUM charts now occupy a broad methodological spectrum. Some adjust thresholds, some estimate alternatives, some update baseline parameters from the outset, some track recurrent episodes without restart, and some embed post-signal diagnosis directly in the chart statistic. The term has therefore become an umbrella for sequential CUSUM procedures that preserve cumulative evidence while relaxing the fixed-model assumptions of the classical chart.