William Gieng
Staff Data Scientist, Qualtrics
Portfolio
Staff-level data scientist focused on causal inference, experimentation, and the translation of statistical evidence into senior leadership decisions.
Case studies translating ambiguous business questions into defensible analytical decisions.
01 — Retention
When a competitor disruption drove a 280K subscriber spike, was it real growth or borrowed users? Three independent lenses on the same question.
Decision impact
Reframed leadership conversation from acquisition volume to activation quality and retained LTV.
02 — Prediction
Can first-month behavior separate "binge and leave" users from habit builders? Acquisition channel, billing, and engagement decomposition.
Decision impact
Surfaced channel-level churn risk and "binge and leave" patterns informing onboarding and acquisition spend.
Long-form essays on causal identification, measurement, and the discipline behind decision-grade analysis.
Separating overlapping churn drivers when promo expiry and use-case completion arrive together.
Connecting LLM summarization failures to causal identification discipline.
Why LLM-extracted variables carry a selection, timing, and measurement footprint the downstream model never sees.
Public notebooks demonstrating the methods behind the case studies and writing.
DiD, synthetic control, RDD, and ITS on a synthetic subscription pricing dataset, with LTV translation and breakeven price calculation.
Three-stage LLM pipeline (extract, synthesize, audit) for decision-grade meeting analysis with explicit grounding and bounded fabrication controls.
Pipeline that extracts and categorizes churn reasons from unstructured text using weak supervision, transformer fine-tuning, and prompt-engineered LLM summaries.
A LinkedIn series on the framing decisions that determine whether analysis is causal.
Parallel trends is not a technical footnote. It is the assumption that makes the comparison believable. When it breaks, the estimate can still look precise while pointing strategy in the wrong direction.
Not every business problem needs an experiment, but every impact claim needs the right identification strategy. Randomized experiment, DiD, ITS, and synthetic control each carry assumptions the data has to earn.
The hardest part is rarely running the model. It is knowing what kind of answer the data can actually support. That is a judgment problem, not a tooling problem.