I’ve sat across from a lot of founders with a lot of forecasts. Three types show up more than any others.
The first is the founder who built a budget in December, opened the file in March when something felt off, and discovered the model had been gathering dust since January 3rd. The assumptions were vague — some percentage growth on last year, a few expense line items, maybe a note about hiring someone in Q2. No metrics. No drivers. Just a spreadsheet that felt like a commitment at the time and became irrelevant by February.
The second went the other direction entirely. Their model pulls from every data feed in the business: cost per impression, clicks per thousand impressions, website sessions, add-to-cart rate, cart-to-checkout rate, checkout-to-order rate, and seventeen more inputs I had to squint to read. It looked like NASA built it. The problem was that when revenue missed by 12%, nobody could tell you why. The signal was buried under so many inputs that the model produced noise instead of insight.
The third is the most painful. It’s the model built by an accountant — precise, beautiful, and completely disconnected from decisions. Every expense line itemized. Every revenue stream mapped to three decimal places. The kind of artifact that takes eight hours a month to update, generates a 40-tab Excel file, and gets quietly ignored by the fourth month because the founder can feel, viscerally, that it adds no decision value. They’re paying for it and not using it. Which means they’re paying for nothing.
Here’s what all three have in common: none of them forecast cash.
They forecast a P&L. Maybe they tack on a line for debt service. And they call it a day.
That’s not a model. That’s a reporting artifact with a budget column stapled to it.
Every decision in your business is a capital allocation decision. Every single one. When you hire someone, you’re allocating capital. When you run a campaign, you’re allocating capital. When you take on inventory, you’re allocating capital. Even your own time — your labor hours — is capital being deployed somewhere.
If your model doesn’t tell you what cash looks like at the end of the month, three months out, six months out, it cannot support decisions. It can report on them after the fact. That’s a different, considerably less useful thing.
This is Lever Four. We’ve covered the Sales Engine (Lever One) — how revenue is built from CAC, LTV, and predictable acquisition. Then Working Capital (Lever Two) — how cash moves through the operating cycle and how working capital days determine whether growth feeds or starves the business. Then the Margin Machine (Lever Three) — what actually falls to the bottom line after you account for the true cost of a sale, channel by channel, SKU by SKU.
This article wires all three together into a single operating model. One that tells you not just what you earned, but what you can actually do with it.
The Right Number of Assumptions
Before we build anything, let’s talk about what to resist.
I took over a client’s model that tracked every step of their acquisition funnel: cost per impression, clicks per thousand impressions, sessions, add-to-cart rate, cart-to-checkout, checkout-to-order. All legitimate metrics. All valuable in the right context. All completely wrong as primary drivers of a financial model.
The problem isn’t that those metrics don’t matter — they do. The problem is that when you route all of them into a model as live assumptions, the model can no longer tell you what happened. A revenue miss gets absorbed across six variables simultaneously. Was it impressions? Was it add-to-cart? Was it checkout abandonment? You cannot see it. The model generates false precision that buries the signal in noise.
We simplified it to two numbers: Cost per Order and Marketing Spend. That’s it.
From those two inputs you get customers acquired. You already know AOV. From AOV and your gross margin (Lever Three), you know what the order produces before fixed overhead. That’s the model.
The underlying funnel metrics didn’t disappear — they still matter. But they belong in a separate analysis layer, pulled out when you’re forming a hypothesis or diagnosing a specific problem. On a weekly basis, track two numbers to confirm you’re trending in the right direction. On a monthly or quarterly basis, go under the hood to understand what’s driving them and where to focus next.
The financial model that analyzes each funnel step in its own cell isn’t being thorough. It’s making the important things harder to see.
Building the Model: Three Statements, Not One
A model that serves decisions has to be a three-statement model: P&L, balance sheet, and cash flow statement. Not because accounting requires it. Because cash flow is where every decision ultimately lands.
Here’s the minimum viable structure for an e-commerce company:
Revenue layer:
New Customers = Marketing Spend ÷ CAC
Returning Customers = Prior Customer Base × Reorder Rate
Total Orders = New Customers + Returning Customers
Revenue = Total Orders × AOVGross margin layer:
Gross Margin per Order = AOV
- Product / Materials Cost (feeds inventory balance)
- Payment Processing (~2.5–3.5% of AOV)
- Shipping Cost
─────────────────────────────
= Gross Margin per Order
Total Gross Margin = Gross Margin per Order × Total OrdersThis is the structure we laid out in Lever Three. If you haven’t done that SKU-level analysis yet, go back. You need a real number here — otherwise the model will tell you a story that isn’t true. Note that CAC sits in a separate line in your P&L as a marketing expense; it’s not baked into gross margin. Both matter, but they live in different places in the model.
Below the line:
Subtract fixed operating expenses — team, software, rent, anything that doesn’t move proportionally with volume. What remains is EBITDA.
Then get into capitalization. Debt? Model the interest expense, principal payments, and draw activity. Equity investors? Model distributions or capital calls. This is also where the working capital mechanics from Lever Two live — AP days, inventory days, and AR days all flow through the balance sheet, which then informs the cash flow statement and shows you what the bank balance actually looks like at the end of each month.
A few key metrics synthesize all of this into numbers you can track on a regular basis:
CAC — cost to acquire a new customer
AOV — average order value (or cart size)
Gross Margin % — what percentage of revenue survives after product, fulfillment, and payment costs
New Customer Count — the acquisition volume driver
Reorder Rate — the retention driver (more on this in Lever Nine)
Working Capital Days — DIO + DSO − DPO, the cash cycle timing
Ending Cash Balance — the number everything else feeds
Track those seven metrics and you’re running the business. Everything else is either a sub-driver you pull when you need to diagnose something, or a rounding error you average and move past.
Rolling Forecasts and the Feedback Loop
Here’s what a static annual budget actually does: it gives you a document to feel good about in December and quietly abandon by March.
The shift to a rolling forecast isn’t primarily about mechanics — updating monthly, extending the horizon as you go. It’s about what the conversation becomes.
Instead of “are we on budget?”, the question becomes: were we right about that experiment?
That reframe changes everything. Variance isn’t a performance score anymore. It’s a signal about whether your assumptions were correct. When EBITDA comes in five points below plan because CAC ran 30% hot, that tells you something specific: either there’s a systemic issue in how you’re acquiring customers, or the market shifted and your targeting hasn’t responded. The variance isn’t a failure. It’s a diagnostic.
And if you’re running the business this way — treating each initiative as a hypothesis, collecting the data, operationalizing what worked, and moving on — the model lets you do that faster. Form the hypothesis. Set the assumption in the model. Run the experiment. Check the variance. Take the lesson. Update the assumption or fix the system. That’s the scientific method applied to a business, and the rolling forecast is what makes the feedback loop tight enough to actually be useful. The goal isn’t to run fewer experiments. It’s to fail faster on the ones that don’t work and double down on the ones that do, as quickly as the data collection allows you confidence.
The most common place this loop breaks is at the data input stage. Models built on hourly billing don’t get updated frequently because frequent updates are expensive and the founder stops owning it. It becomes something that happens to them rather than something they use. I build on templated QBO integrations specifically to make actuals fast to load — because speed creates ownership, and ownership creates the habit of actually looking at what the variance is telling you.
When you do look at variance, the question is always: why were we wrong? The honest answer usually points to one of two things — an incorrect assumption about how the business works, or a missing or broken system. CAC running 30% above plan for three consecutive months doesn’t mean your marketing got unlucky. It usually means nothing is reacting to what the market is telling you. That’s more valuable to surface than the number itself. The number is a symptom. The system gap is the diagnosis.
Scenario Planning: Making Capital Allocation Visible
The most underused part of a financial model isn’t the base case. It’s the scenario layer.
A client of mine sold high-ticket products with strong gross margins — the COGS as a percentage of AOV was the same whether the cart came in at $200 or $180. Same product mix, same fulfillment cost ratio. The difference is what happens to that $20. Because CAC is a flat cost per customer — you spent $X to acquire them regardless of what they put in the cart — that $20 incremental AOV isn’t burdened by any additional acquisition cost. It falls almost entirely to the bottom line. We built a scenario around it: what happens if we run out of the specific products that tend to pull cart size toward $200? The answer was ugly enough that inventory management on those SKUs became a hard operational priority. Not because we made a sophisticated argument. Because the model made the downside impossible to dismiss.
The other application I use constantly is headcount decisions, and the model makes them considerably less gut-driven.
Say you’re evaluating a marketing director. The thesis: they improve CAC from $25 to $20. Realistic assumption: two months to learn the business, five more months to execute the improvement fully. You model that timeline against the cash outlay during their ramp, layer in the incremental gross margin once the improvement lands, and check the ending cash balance at every point along the way.
Now extend that logic one level. A lower CAC means more customers acquired per dollar of marketing spend — which at constant AOV directly lifts revenue. If that same marketing director also improves website conversion by 1%, on a $3M business that might be $50–60K in additional gross margin per year, without spending another dollar on acquisition. Those two levers are connected: CAC improvement and conversion improvement are often driven by the same person executing the same strategy. The model lets you quantify both, see the combined cash impact across a realistic ramp timeline, and answer the actual question — not “can we afford to hire?” but “what is this hire worth, over what timeline, and does our cashflow runway support the risk?”
Same logic applies to a sales rep. Revenue per rep is a starting point, not the answer. What’s realistic quota attainment in months one through three versus month six? What does the training burden cost in senior team time during ramp? Does adding this rep require incremental inventory to support the volume? What’s the cash trough from start date to meaningful margin contribution — and does the ending balance stay above your minimum comfort threshold throughout? The model tells you to hire when the expected impact is defined, the cashflow runway absorbs the ramp, and you have the operational signals that the system can actually support the volume.
Operational signal first. Cash confirmation second. In that order.
A Note on Inventory
The financial model should not be driving purchase orders. Inventory needs its own model or at minimum its own logic layer — feeding the financial model’s inventory balance and the AP line, not the other way around.
What matters at the financial model level is the cash impact of your inventory assumptions, and there are two conversations worth having here that most founders never get to.
The first is AP terms. What happens if we extend from 30 days to 60 days with our primary supplier? At scale, that change can unlock six figures of working capital that was sitting invisible inside a default term assumption. The answer from the supplier might be no — but you can offer a small margin improvement as a negotiating chip, effectively treating extended terms as a cheap borrowing mechanism. You can only have that conversation if you’ve already run the scenario and know what the number is worth.
The second is inventory days. Getting from 75 days of inventory on hand to 45 days isn’t just an operations win — it’s a capital release. On a business carrying $500K in inventory, that shift frees roughly $200K in cash that was otherwise sitting on a shelf. The model shows you exactly where that capital goes when it’s unlocked, and whether it’s better deployed back into marketing, used to pay down a line of credit, or held as a cash cushion. That’s not an ops conversation. That’s a capital allocation conversation, and only the model can frame it that way.
Common Patterns I See Constantly
Too many models. I’ve walked into situations where the founder had four separate files: the one the accountant built, the one from last year, the one with the new sales strategy baked in, and the one they actually look at each month. None of them talk to each other. Have one model. Maybe two if you’re tracking a genuinely distinct scenario. More than that and the maintenance burden kills the habit of using any of them.
Percentage-based growth assumptions. “We’ll grow 2% per month” is not a forecast. It’s a wish attached to a spreadsheet. A driver-based model tells you which input was wrong when revenue misses. A percentage-based model just tells you that you missed. Drivers. Always drivers.
Rounding errors masquerading as strategic line items. I’ve seen models that break out every individual SaaS subscription for a company doing $800K in revenue. The marginal $200 of software spend won’t make or break any decision at that stage. Contribution margin and the sales engine will. Take an average for the noise, adjust for lumpy payments billed quarterly or annually, and spend your time on the variables that actually move the needle.
Using historical numbers as assumptions without questioning the system. There’s an implicit belief when founders anchor assumptions to history that prior performance reflects a functioning system. Often it doesn’t. If a sales rep has historically closed $400K per year, that’s data — but it’s not a quota if the CRM is a mess, lead quality has never been analyzed, and there’s no structured process underneath it. Build from first principles. What should this rep produce if the system is right? Use that number to identify the gap between where you are and where you’re going.
CEO Q&A
“I built a P&L forecast in December. Isn’t that basically the same thing?”
No. A P&L forecast tells you what you expect to earn. A three-statement model tells you what cash you’ll have when a vendor invoice lands, a large inventory order goes out, and a loan payment hits — simultaneously. Those are the moments that create crises, and they’re invisible in a P&L-only view. The P&L is the input. The cash flow statement is the output that actually runs the business.
“My model has thirty inputs. More inputs means more accuracy, right?”
More inputs means more places for the model to be wrong and less ability to explain any individual miss. If revenue comes in light and you have thirty variables, you have thirty possible explanations and no clear answer. The goal is the minimum number of assumptions that explain the maximum amount of variance. For most e-commerce businesses, five to seven driver metrics gets you there. Start there and add complexity only when a specific variable is genuinely moving the needle and you need visibility into why.
“We’ve been growing around 2% per month. Can I just use that?”
You can, but you won’t learn anything from it when it breaks. What drove that 2%? New customers? Higher AOV? Better retention? If you don’t know, you don’t know which lever to pull when growth stalls. Build from drivers and the 2% becomes an output to pressure-test, not an assumption to start from.
“My accountant updates the model. Why do I need to be involved?”
Think about how many ideas are floating in your head right now. New hires, channel experiments, pricing changes, inventory bets — probably a dozen things you want to test. A good model is how you decide which ones are worth your time. You take the initiatives that feel most impactful, run the scenarios, and find out which one actually moves the needle most. Then you execute, measure the variance, and learn something. Then you do it again. That’s the scientific method applied to a business.
The model is most powerful when it’s a live part of how you think — something you’re in regularly, not something that gets delivered to you. A fractional CFO or finance partner can help you build it, keep it current, and pressure-test your assumptions. But the decisions it informs are yours, and the habit of using it has to live with you. If you’re only seeing the output once a month in a summary email, you’re getting reporting. Reporting tells you what happened. The model tells you what to do next — and that only works if you’re in the room when the scenarios are running.
“The model says I can afford another sales rep. When should I pull the trigger?”
Affordability is necessary but not sufficient. The more useful question is: what is the expected return, and does the cash trough during ramp stay above the floor you’re comfortable with? A rep doing $600K at full productivity sounds straightforward — but what’s realistic in months one through three while they’re learning the product? What does senior team time spent on training actually cost in terms of their own output? Does the incremental volume this rep generates require you to carry more inventory, which hits cash before the revenue arrives? And if your quota attainment assumption is aggressive, what does the scenario look like if they hit 70% of it in year one instead? Model the ramp, model the inventory impact, model the cash trough. When the downside scenario still keeps you above your minimum balance and the operational signals say the system can support the volume, move.
“Why does variance analysis matter if I already know the number came in low?”
Because the number is the symptom. Say gross margin came in three points below plan for the second consecutive month. You could note it and move on. Or you could pull the variance and find out that shipping costs spiked because a fulfillment system change created exceptions that nobody caught. That’s not a margin problem — that’s a missing system. The variance pointed you there. Without it, you’re trimming your growth assumptions to account for a problem that could be solved with a process fix. The diagnostic is more valuable than the score.
What Comes Next
The next lever is The People Math — every hire as an investment with a return profile, how to know when demand actually justifies the headcount, and how the operating model we built here makes the hiring decision a calculation instead of a leap of faith.
If you’re not subscribed to Conduit of Value, now is a good time. The series is designed to build — each lever sharpens the ones before it, and this one was the connective tissue.
And if any of this resonated because you recognized your own model in the opening, reach out. I build these for a living, and the first conversation is always free.
Lever One: The Sales Engine | Lever Two: Working Capital | Lever Three: The Margin Machine