Forecasting the Cash That Isn't From Customers
Customer payments were the largest part of incoming cash, covered in Parts 1 and 2. But they were not the whole picture. I checked ten other possible sources of incoming cash, one at a time, against real history. Seven turned out to be zero. Three needed a real method — and one of them, tax refunds, needed the most work: the first idea for how to calculate it was wrong, and had to be replaced with one tested against real numbers.
Going through every category
Before forecasting any of this, I went through a fixed list of possible non-customer income sources. For each one, I checked two things: did this actually happen in the last 12 months, and is it planned in the next 3 months? Where the answer to both was no, the category got excluded from the forecast.
If an excluded category receives cash, I check whether it is a one-off or likely to happen again. If it is likely, I include it in future forecasts. Otherwise, I keep it excluded.
| Category | Found in the last 12 months? | What I did with it |
|---|---|---|
| New loans or owner investment | No | Not included |
| Selling equipment, property, or other assets | No | Not included |
| Tax refunds (VAT) | Yes | Built a real forecasting method — covered below |
| Refunds and rebates from suppliers | Yes, but small | Too small on its own to require a separate forecast line |
| Bank account cashback | Yes | Modeled as a percentage of forecasted revenue — covered below |
| Grants and subsidies | Yes, 3 payments | Tracked only when a specific application is pending — covered below |
| Legal settlements | No | Not included |
| Insurance payouts | No | Not included |
| Renting out unused space | No | Not included |
| Employee repayments | Negligible | Not included |
Seven of these ten categories are excluded, and that is not a gap in the forecast. Each one got checked against real data and real management plans before being ruled out.
Cashback and grants: same size, opposite methods
Two questions decide how to forecast any small, non-customer income source:
Is it large enough to be meaningful? If a source is a small share of total forecasted cash, spending heavy effort on it does not add enough value to justify the time. The amount of effort should match the size of the number.
Does it happen continuously, or as separate events? Some income adds up a little at a time — like interest or cashback. Some income arrives as a small number of separate events, decided by someone else — like a grant. These two shapes need opposite methods, even when the two numbers are about the same size.
Cashback and grants are a clear example of this. Both worked out to a similar size: roughly 0.15% of total sales for cashback, roughly 0.13% for grants. Same size. Completely different treatment.
Cashback shows up once a month, in the first week of the month, not every week. I checked the real payment history to see when it actually arrives. It always arrives in the first week of a new month, never spread across the other weeks. So the forecast only puts a number in that one week, and leaves every other week at zero for this line.
The amount uses a simple calculation: add up all incoming money from the previous month, then take 0.15% of that total. That 0.15% number comes from checking history — actual cashback received, divided by actual total incoming cash, over many past months.
This timing is specific to this bank. A different bank may pay cashback more often, on a different day, or calculate it a different way entirely. Before using this method, the real pattern needs to be checked in the actual bank statements — not assumed to be the same everywhere.
Grants happen as separate events. Only 3 grant payments arrived in 12 months. Three events are not enough to calculate a real average — any average from 3 numbers is really just those 3 numbers. It can look like a pattern, but there is not enough data for a real pattern to exist. Turning that into "a small amount every week" would make a rare, uncertain kind of income look steady and guaranteed, which it is not.
Not as an average. If a grant application is currently submitted and waiting on a decision, it gets logged directly: the amount, the expected decision date, and a realistic percentage chance of approval. If nothing is currently pending, nothing gets forecast for grants at all. An unexpected grant then becomes a good surprise, instead of a number the forecast promised and did not deliver.
VAT refunds: a real pattern worth modeling
Tax refunds turned out to be different from both cashback and grants. They happen often enough, and predictably enough, to deserve a real formula — but only after the first idea for one turned out to be wrong.
The first idea was simple: higher sales should lead to higher VAT refunds. That seemed logical, so I tried tying the refund estimate directly to forecasted sales. It turned out to be wrong. Higher sales does not directly cause a higher VAT refund. Higher purchases does — VAT refunds come from VAT paid on business purchases, not from sales.
So the method changed. Instead of sales, I tied the estimate to forecasted VAT taxed purchases and expenses.
First, from 12 months of real history: total VAT refunds received, divided by total VAT-eligible purchases in that same period. That gives one ratio — for example, a certain number of cents refunded for every euro spent on eligible purchases. Then, going forward: forecasted purchases and expenses for a future month, multiplied by that same ratio, gives the projected VAT refund for that month.
One more pattern showed up while checking the history: VAT refunds mostly arrive within the first 10 days of a month, not spread evenly across the month. That timing gets used too — not just the total amount, but roughly which week of the month to expect it in.
This is worth comparing to grants directly, since both are technically "money from the government." Grants are 3 events in a year, with no pattern to calculate from — they get tracked as a pipeline, one at a time. VAT refunds happen roughly every month, tied to a specific, known cause — purchases — which makes a real formula possible. Same broad category of income. The amount of real history available for each one is completely different. So the methods are completely different too.
Why this matters to a client
A forecast that tries to model everything with equal effort wastes time on things that barely matter, and risks rushing through the things that do. Matching the method to how each income source actually behaves — instead of using one formula for everything — is what keeps a forecast both accurate now and realistic to keep updating later.
This is Part 4 of a 5-part case study. Part 5 — comparing the forecast against what actually happened — publishes soon.