LatentView & Affine: Analytics / Data Science

import pandas as pd, numpy as np
df = pd.read_csv("campaign.csv")
df.columns = df.columns.str.strip().str.lower().str.replace(" ", "_")     # tidy names
df = df.drop_duplicates()
df["date"] = pd.to_datetime(df["date"], errors="coerce", dayfirst=True)   # fix types
df["spend"] = pd.to_numeric(df["spend"].astype(str).str.replace(r"[₹,]", "", regex=True), errors="coerce")
df["channel"] = df["channel"].str.strip().str.title().replace({"Fb": "Facebook"})
df["spend_missing"] = df["spend"].isna().astype(int); df["spend"] = df["spend"].fillna(df.groupby("channel")["spend"].transform("median"))
q1, q3 = df["spend"].quantile([.25, .75]); df["spend"] = df["spend"].clip(upper=q3 + 1.5 * (q3 - q1))
df = df.dropna(subset=["date", "campaign_id"])

• Explain each step: names, duplicates, types, standardised categories, missing values (with a flag), outliers, required keys. Then validate with df.info() and describe().

SELECT department,
       AVG(CASE WHEN salary > 2000000 THEN 1.0 ELSE 0 END) AS high_earner_share,
       RANK() OVER (ORDER BY AVG(CASE WHEN salary > 2000000 THEN 1.0 ELSE 0 END) DESC) AS rk
FROM employees
GROUP BY department
HAVING COUNT(*) >= 10;          -- avoid tiny departments dominating

• Incrementality first: use a holdout or geo test, or a matched control, to estimate the incremental conversions and revenue caused by the campaign.
• ROI = (incremental gross margin − campaign cost) / campaign cost.
• Attribution for always-on channels: marketing-mix modelling (MMM) with adstock and saturation, or data-driven multi-touch attribution where tracking allows.
• Report confidence intervals and cost per incremental acquisition, and recommend a budget reallocation.

• Visual: pair plots or PCA/t-SNE for intuition.
• Formal: fit a linear SVM with a very large C (a hard margin) or a perceptron. Zero training error means linearly separable. Linear programming feasibility is the exact test.
• If not separable: kernels (RBF), feature crosses, trees or neural networks.

• Scope tightly: agree the 2–3 key questions and the deliverable format on day 1.
• Plan: a work breakdown with owners, daily stand-ups, risks (data access) escalated early.
• Quality: peer reviews of SQL and logic, sanity checks against known totals.
• Communicate: a mid-point preview to the client; then the final story with insight, recommendation and impact.
• Show delivery + people management + client handling.

• Pre-register the metric, the minimum detectable effect, α and power, and the duration.
• Check for sample-ratio mismatch; compute the lift with a confidence interval (z-test or bootstrap).
• Significant and practically meaningful (above the MDE) and guardrails intact → ship. Otherwise iterate.
• Watch for peeking, novelty effects, and multiple comparisons.