Schedule Credibility under Capacity-Constrained Attention: Multi-Source Evidence from High-Mix Semiconductor Equipment Manufacturing
Year: 2026
Type: preprint
Abstract: A production target is useful only while downstream decisions can still act on it. This paper investigates schedule credibility in high-mix, low-volume (HMLV) semiconductor equipment assembly using an empirical archive of 53,860 production snapshots covering 562 units and 111 planning dates, integrated with independent capacity/headcount models and material-shortage tracking registers. Across 307 units with traceable milestone histories, 124 verified target-date revision events are identified. Rather than treating operational records as clean administrative facts, measurement uncertainty is modeled as a primary operational constraint. A hierarchical beta-binomial model partially pools family-level revision incidence, while a pattern-mixture analysis shows that plant-wide estimates remain bounded within a sensitivity region of 20.7% to 30.8%. Recurrent-event Prentice–Williams–Peterson (PWP) gap-time models, stratified across sequential revision orders and validated through Schoenfeld residual diagnostics, separate structural product vulnerabilities from transient operational shocks. Revisions executed strictly after the superseded due date (49.2% of events) exhibit a median displacement of 10 days, compared to 2 days for proactive adjustments (Cliff's δ=0.642). Exact tests via Monte Carlo simulation confirm that revision logging clusters heavily by calendar weekday (p<10 -6 ), while independent capacity records demonstrate that aggregate labor-hour slack coexists with peak concurrency overload. Machine learning models predicting next-update target adjustments achieve strong discrimination (AUROC 0.861–0.916) but low absolute precision (AUPRC 0.0095–0.0380) due to extreme positive-class rarity. Decision Curve Analysis confirms that ranking by temporal probability delivers strictly positive net benefit over default heuristics across operationally relevant intervention thresholds. A capacity-constrained decision model demonstrates why complex prioritization rules do not dominate simple temporal heuristics under tight review bandwidth: at a 10% review budget, static due-date proximity captures 79.0% of revision events and 72.1% of revision-day burden, statistically dominating complex multi-factor blends. We translate these empirical findings into a dual-timescale Lean Six Sigma control architecture: a fast attention loop targeting imminent temporal vulnerability, coupled with a slow improvement loop resolving recurrent structural root causes. Competing mechanisms—information latency, administrative batching, concurrency bottlenecks, and strategic withholding—are formalized within a Causal Directed Acyclic Graph (DAG) with explicit identification boundaries.
Source: SSRN Electronic Journal
DOI: 10.2139/ssrn.7536343
Author: Ngoc Huy Mai
Language: English
Subfield: Strategy and Management
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7212181237