Data Science and Operations: How the Pieces Fit Together
The four levels of analytics
Most data work falls on a ladder, and the ladder runs from describing what already happened to recommending what to do next. Descriptive analytics answers “what happened?” Cohort retention tables show how each signup month survives across periods. Conversion funnels show where users drop off between landing and purchase. There's no model under the hood, just well-organized counts. The Cohort Retention Table and the Conversion Funnel Drop-Off Analyzer sit at this level.
Diagnostic analytics asks “why did it happen?” Correlations and correlation matrices flag which variables move together, which is the first step toward identifying drivers. The Correlation Calculator and Correlation Matrix Visualizer live here.
Predictive analytics answers “what's likely to happen?” Logistic regression predicts conversion probability from input features. Time-series decomposition separates trend, seasonality, and noise so you can extrapolate. Monte Carlo simulation generates a probability distribution over future outcomes when inputs are uncertain. The Logistic Regression Probability Curve Visualizer, Time Series Decomposition Demo, and the Monte Carlo Simulator cover this band.
Prescriptive analytics asks “what should we do?” Linear programming maximizes profit subject to resource constraints. EOQ tells you the order quantity that minimizes inventory cost. Six Sigma turns defect counts into capability indices that drive process changes. The Linear Programming Solver, EOQ Calculator, Safety Stock Calculator, and Six Sigma Calculator sit at the top of the ladder.
The honest version: most teams spend 80% of their time at the descriptive and diagnostic levels, 15% on predictive, and almost none on prescriptive. Operations research lives at the prescriptive end and is its own discipline.
Frequentist vs Bayesian: when each wins
The frequentist tradition (which the A/B Test Significance & Lift Calculator uses) treats the population parameter as fixed and asks: given this data, how surprising would the result be if there were really no effect? p-values and confidence intervals come from that framing. It's well-tested, well-tooled, and what most published research uses.
The Bayesian tradition treats the parameter as a probability distribution and asks: given the data and a prior belief, what's the updated probability that the effect is positive? Bayesian A/B testing has gained ground in industry over the past decade because the output (“there's an 87% probability that variant B beats variant A”) is easier to explain to product managers than a p-value. Frequentist tools are stronger when you have no useful prior, when peer reviewers expect classical tests, or when sample sizes are large enough that the prior gets overwhelmed by data anyway. Bayesian tools win when priors are informative (you've run hundreds of similar tests), when sequential testing matters (peeking is fine in proper Bayesian frameworks), or when the audience needs probabilities rather than p-values. Plenty of teams run both and report whichever lands more cleanly.
Statistical significance vs practical significance
The single most-misused concept in analytics. A test can be statistically significant and operationally pointless at the same time, because statistical significance only tells you the effect is unlikely to be zero, not that the effect is large enough to matter.
Take an A/B test on a checkout button color. With 500,000 visitors per arm, you can detect a 0.3% lift with p < 0.001. The result is statistically significant and substantively trivial. A 0.3% lift on a $40 average order value moves $0.12 per visitor. If the engineering effort to ship the change costs more than the lift recovers, you've discovered a real but worthless effect. The Sample Size & Power Calculator tries to head this off by asking you to specify a Minimum Detectable Effect (MDE) before launching: the smallest lift you'd actually act on. Cohen's effect size framework formalizes the question: for means, d under 0.2 is small, around 0.5 is medium, 0.8 is large. For correlations, r below 0.3 is weak. These are heuristics, not laws, but they help anchor whether a “significant” result is worth shipping. Significance answers “is this real?” Effect size answers “is this big?”
Where operations research fits among data tools
Most data content stops at statistics and machine learning. Operations research is the sister discipline that picks up where prediction ends and asks “given what we know, what should we actually do?” The split shows up in the calculator stack here.
Statistics tells you whether a treatment works. The A/B Test Calculator computes a p-value for whether the lift you observed is likely real. ML predicts an outcome from features; the Logistic Regression Visualizer shows the sigmoid. Neither tells you how many call-center agents to staff at 2 PM on a Tuesday in November.
The Queueing Calculator does. M/M/c gives you P(wait ≤ t) and Wq directly from arrival and service rates. There's no statistical inference happening, just a closed-form model of how queues behave at steady state. Same story with the EOQ Calculator, which solves a constrained-optimization problem to minimize total inventory cost. EOQ has nothing to do with hypothesis testing. It's calculus on a cost function. Six Sigma sits even further from inference. Process capability indices like Cp and Cpk are deterministic transforms of process variation against spec width. The point isn't to test whether your process is in control. It's to convert defect counts into a number that triggers a specific action: improve variation, recenter the process, or accept the loss. Once you're past “is this real?” and onto “what do we do about it?”, you're in OR territory.
When you need a statistician (not a calculator)
A calculator like the ones on this site works for exploratory analysis, sanity checks, and preliminary planning. For some decisions, that's genuinely insufficient and it's worth being explicit about which.
Clinical trial design is the cleanest case. The FDA, EMA, and equivalent regulators expect pre-registered protocols with documented power calculations, interim-analysis plans, multiplicity adjustments, and stratification schemes. No web calculator generates the documentation regulators require. Hire a biostatistician.
Regulatory filings (FDA, SEC, FCC) bring similar constraints. Anything that ends up as evidence in a financial disclosure or a drug-approval submission needs a documented statistical workflow with audit trails, not a calculator output someone took a screenshot of. Causal inference for legal discovery and damages estimation is its own deep specialty. Wrongful-termination disparate-impact cases, antitrust price-fixing models, lost-profits calculations: those get challenged on cross-examination and need a statistician who can defend the methodology in court. Bivariate correlation will not survive that examination.
Material business decisions that hinge on whether a treatment effect is real also belong in this bucket. Pricing changes that move millions in annual revenue, product launches with multi-year roadmap implications, M&A diligence on subscriber LTV. The calculators on this page can frame the analysis, but the final number deserves a domain statistician's review and an experimental design built specifically for that decision. The honest framing: this site exists for the 90% of analytics work that's exploratory, educational, or preliminary. The 10% that's high-stakes deserves more than a calculator and a coffee break.
The calculator stack: pick the right tool
A short decision guide, organized by what you're actually trying to do.
Analyst trying to understand data. Start with the Correlation Calculator for two variables, the Correlation Matrix Visualizer for many. Use the Cohort Retention Table to see how patterns differ across groups. The Time Series Decomposition Demo separates trend from seasonality if your data has a time dimension.
Analyst running an experiment. Plan with the Sample Size & Power Calculator before launch. Analyze with the A/B Test Significance & Lift Calculator once you have data. Evaluate classification models with the Confusion Matrix Calculator. For feature engineering ahead of any ML model, the Feature Scaling & Normalization Helper tells you which scaler to pick and how to avoid leakage.
PM modeling unit economics. The CAC, LTV & LTV/CAC Calculator gives you the snapshot. The CLV Scenario Simulator lets you compare retention scenarios. The Cohort Retention Table shows how those numbers actually evolve. The Conversion Funnel Drop-Off Analyzer finds the leakiest stage to fix first. The Basic Churn & Retention Calculator is the simplest entry point if you're new to subscription metrics.
Ops manager planning capacity. Start with the Erlang C & M/M/c Queueing Calculator for general sizing, or the Queue Wait-Time SLA Calculator when you have a contractual percentile target. The Safety Stock & Reorder Point Calculator and EOQ Calculator handle inventory.
Project manager pricing risk. The Monte Carlo Simulator generates outcome distributions for any model with uncertain inputs. The Project Monte Carlo Risk Calculator specializes in three-point estimates and critical-path schedule risk. Use the ROI / NPV / IRR Calculator once you have a cash-flow model to evaluate.
Engineer doing process control. The Six Sigma Calculator converts defect counts into DPMO, sigma level, and Cpk. Pair it with the Time Series Decomposition Demo when you need to detect trend or seasonality in process data. The Markov Chain Steady State Demo helps when transitions between discrete states matter.
If you're between tools, the glossary below explains the relationships.