Building a startup financial model

Founders build models for three different audiences and often conflate them. The operating model answers what to do this quarter: can we afford this hire, what happens to runway if we increase marketing. The fundraising model shows an investor the shape of the opportunity. The board model tracks actual results against the plan.

They share a structure but not a purpose. The operating model should be pessimistic because decisions are made from it. The fundraising model should be ambitious but defensible. Using one for the other's job is how companies hire against revenue that never arrives.

A weak model types a revenue number into each month and grows it by a percentage. A driver-based model builds revenue from the things that actually produce it — marketing spend produces leads at a cost per lead, leads convert at a rate, customers pay a price and churn at a rate — so that changing an input changes the output through a chain you can inspect.

The value is not accuracy. It is that when reality diverges from plan, a driver model tells you which assumption was wrong. A model with revenue typed in tells you only that you missed.

A complete model produces a profit and loss statement, a balance sheet and a cash flow statement, and they tie together. The reason to bother with all three, rather than just the P&L, is that profit and cash are different and companies die of the second.

Annual contracts paid upfront generate cash long before the revenue is recognised. Customers on ninety-day payment terms generate revenue long before the cash arrives. A business can be profitable on paper and insolvent in practice, and only the cash statement shows it.

In almost every software model, three inputs move the outcome more than everything else combined: churn rate, conversion rate and average revenue per account. Small changes compound over the forecast period, and a model that is wrong about churn by one percentage point per month is wrong about the three-year revenue figure by a very large margin.

The practical move is to test sensitivity explicitly. Run the model at your base case, then at plus and minus twenty percent on each key driver, and see which one breaks it. That tells you what to measure obsessively in the business.

Most models fail obvious tests. Check that headcount grows in a way a human could actually hire — thirty engineers in one quarter is not a plan. Check that revenue per employee lands in a believable range for your model. Check that the implied market share at the end of the forecast is not a multiple of the market's actual size.

The most common failure of all is the hockey stick with no mechanism: growth that accelerates in month eighteen because the model needs it to, with no corresponding change in spend, headcount or conversion to cause it.

Building too much detail is the classic waste. A forty-tab model with individual line items for software subscriptions is not more accurate than a clean one-tab model; it is just harder to change, which means it stops being updated.

The second is never comparing forecast to actual. A model that is not reviewed against reality each month is a document rather than a tool, and its assumptions drift from the business without anyone noticing.

Garage to IPO is itself a driver model. Marketing spend produces users at a channel-specific cost, users convert and churn at rates influenced by product quality, revenue is users multiplied by average revenue, and valuation is a multiple applied to that revenue and adjusted by growth.

Playing with the levers is a fast way to build intuition for sensitivity. Cutting churn by a point does more for the three-year outcome than a large increase in marketing spend, and the dashboard makes that visible in a way a spreadsheet rarely does.