SCOPE aligns interventions across multiple decision points using causal learners plus backward induction, consistently beating existing sequential prescriptive process monitoring methods on SimBank and the new SimBPIC17 benchmark

SCOPE embeds causal learners inside backward induction, recursively estimating from the last decision point back to the first the effect of each candidate intervention action on the target KPI, yielding sequential intervention policies aligned across decision points. Existing prescriptive process monitoring methods either handle a single intervention scenario, optimize each decision point in isolation, or rely on MDP or data augmentation to train a reinforcement learning agent; SCOPE works directly on observational event logs and needs neither a process-specific simulator nor log augmentation. The paper provides Q-function and value-function definitions for backward induction, a regret-based value update, estimation procedures for the S-, T-, and RA-learners, and training and inference pseudocode for SCOPE-S; the method is identifiable under three standard assumptions: sequential ignorability, SUTVA, and positivity.

On SimBank, SCOPE outperforms the baselines across nearly all combinations of confounding level and training size, with the only exception being the 1K training set at 0.99 confounding. No prior method combined multi-interventional support, alignment across decision points, and freedom from process approximations; the experiment separates these three properties and shows that aligning interventions is generally more effective than optimizing them independently. Experiments cover confounding levels 0.9-0.99 and training sizes 1K, 10K, and 50K, with 10 random seeds per setting and shaded bands showing one standard error; the random policy yields a gain of -132.597 +/- 0.983 and the upper bound is 411.537.

SCOPE's advantage holds across learner types and base models, and widens as the number of decision points grows. The paper compares S-, T-, and RA-learners with XGBoost, Random Forest, and MLP/LSTM base models on SimBank, and varies the number of decision points from 2 to 6 on SimBPIC17, testing whether the advantage is specific to one configuration. S- and RA-learners perform strongest while the T-learner degrades most under high confounding; XGBoost performs best overall and Random Forest worst; on SimBPIC17 SCOPE's advantage grows with more decision points, though both SCOPE and SEP-S drift further from the upper bound as decision points increase.

The paper releases SimBPIC17, a semi-synthetic benchmark based on the real-life BPIC17_W event log that makes it easy to vary the number of decision points and the confounding level, as a reusable evaluation resource for sequential prescriptive process monitoring. SimBank was previously described as the only simulator designed specifically for prescriptive process monitoring; SimBPIC17 adds a semi-synthetic setting derived from a real log with an adjustable number of decision points. SimBPIC17 introduces an intervention on the existing activity call_incomplete_files, with KPI costfiles = costtpt * tptfiles + n_calls * costcall; the effect of a call depends on loan type and the average duration of previous activities, and the bank policy calls when the loan type is car or loan takeover and the average activity duration exceeds 4025 units; code, algorithms, simulators, and additional results are publicly available in a GitHub repository.

The result targets business processes that take observational event logs as input, contain multiple intervenable decision points, and have a KPI measurable at case end, such as choosing a procedure and setting an interest rate in loan approval, or deciding whether to call a client about incomplete files. It applies when sequential ignorability, SUTVA, and positivity hold and when intervention actions map to existing log attributes. For practitioners, the paper recommends preferring S- or RA-learners with a base model such as XGBoost, since the T-learner degrades markedly under high confounding; SCOPE's gains are clearer with more training data, because error propagation through backward induction weakens as data grows.

SCOPE assumes no interference between cases, and sequential ignorability may not always hold in business processes; the paper lists both as common assumptions and notes that experiments still identify effective policies under realistic confounding levels. The method depends on the predictive accuracy of underlying models, and a highly unpredictable KPI can cause errors to accumulate across decision points, a challenge the paper says affects all methods. In addition, both SCOPE and SEP-S drift further from the exhaustive upper bound as decision points increase, the former from cross-model error accumulation and the latter from treating decision points independently. Current validation rests on two loan scenarios, SimBank and SimBPIC17, and real event logs provide no counterfactual ground truth, so real-world deployment performance remains an open question; the paper also lists handling interference, relaxing sequential ignorability, and reducing computational cost via an end-to-end neural framework as future directions.