CW-201b · Module 2

Variance Analysis & Projections

4 min read

Variance analysis is where Claude earns its keep in financial workflows. The task is inherently tedious for humans — comparing dozens or hundreds of line items between actual and budgeted figures, calculating the variance, determining whether the variance is favorable or unfavorable, and writing commentary explaining the material variances. A human analyst doing this for a 50-line-item budget takes hours. Claude does it in minutes, and it never mislabels favorable as unfavorable.

The prompt structure matters. "Compare actual Q1 results to the Q1 budget. For each line item: calculate the dollar variance and percentage variance. Flag any variance exceeding 10% or $50,000. For each flagged item, provide a one-sentence explanation of the likely driver based on the data available." That prompt produces a structured variance report. "Compare these two spreadsheets" produces a narrative that may or may not cover what matters.

Projections are where you must be most careful. Claude can project future periods based on historical trends — and it will do so confidently regardless of whether the projection is statistically valid. A three-quarter trend does not necessarily predict the fourth quarter. A revenue growth rate calculated from two data points is not a reliable basis for a five-year projection.

The guardrail for projections is mandatory assumption documentation. Every projection Claude produces must include: the data it based the projection on, the method it used (linear trend, growth rate, moving average), the assumptions it made (seasonality, growth rate stability, market conditions), and a confidence range. A projection that says "Q4 revenue: $2.3M" is less useful than one that says "Q4 revenue: $2.1M-$2.5M, based on 3-quarter linear trend, assuming no seasonality adjustment, moderate confidence given limited data points."

Do This

  • Specify variance thresholds explicitly: "Flag variances over 10% or $50K"
  • Require assumption documentation on every projection — method, data, confidence range
  • Use Claude for mechanical variance calculations — it is faster and more reliable than manual work
  • Cross-check Claude's calculations on 3-5 line items to verify the analysis is sound

Avoid This

  • Ask Claude to "project next year's revenue" without specifying the method or assumptions
  • Present a Claude projection as a single number without a confidence range
  • Skip the cross-check — Claude can miscalculate if the data parsing was imperfect
  • Use Claude's projections as the basis for commitments — they are conversation starters, not commitments