Why comparing against the current estimate is misleading
The usual way to measure a forecast is to compare the month's actuals with what the forecast says. The trouble is that the forecast gets corrected with every load: by the time the month ends, the estimate already looks a lot like what happened, and the forecast seems close to perfect.
The useful question is a different one: what was estimated for that month one, three or six months earlier, and how far did reality land from it? Answering it means keeping what was estimated at each point in time.
A snapshot of the forecast every month
FLIQO freezes a snapshot of the forecast every month, on the first day the load reaches the planning horizon the company declared. That way the same month gets estimated many times before it happens, from the moment it enters the horizon until it arrives.
Snapshots don't change. Reprocessing a load or fixing a figure never rewrites one, and a snapshot is only replaced when a person deliberately does so. To start with some history, you can freeze snapshots for past months from the estimates you had at the time.
How far off it is, by lead time
The horizon summary shows how far off the forecast is at one month, three or six, averaged over every month that has already closed. It shows the error against plan, the bias (whether it tends to run short or long), how much that error varies, and how many observations it's based on.
For a specific month, the convergence curve shows how the estimate moved toward the actual as the month got closer.
Receipts and payments, each on its own
Running short on a payment doesn't have the same effect as running short on a receipt, so they're measured separately. FLIQO rates each variance by its effect on cash: conservative if it left more cash than planned, deficient if it left less.
Conservative isn't free either: overestimating ties up liquidity and can lead to funding a shortfall that never existed. That's why only results within the tolerance the company sets show green, with one margin for receipts, one for payments and one for balances.
Error or timing?
Not every variance is a miscalculation. FLIQO splits each month's difference into four parts:
- Timing: the flow was estimated correctly but arrived in the next or previous month. The money wasn't lost, it just shifted.
- Amount: the flow arrived in the expected month, but for a different figure. This is the error worth fixing.
- Unplanned: a flow arrived that nobody had estimated.
- Didn't happen: a flow was estimated and never occurred.
The same error percentage reads very differently when most of it is timing rather than amount. The split is shown in total and by category.
Down to the cause
Categories with the largest error come first, with the amount and percentage of the variance, and each one opens by department to show where it was concentrated. Each shows the cause recorded for that month; if nobody recorded one, a link takes you to the forecast-versus-actual analysis to add it.
Persistent bias alert
If a category is under- or overestimated in the same direction for three months in a row, above the threshold the company sets, the head of the responsible department gets a notice linking to this screen. The alert is optional and doesn't change the forecast: a person decides the adjustment.
Closing balance accuracy too
Flow accuracy has a counterpart for the balance. For a given closing date, FLIQO shows the balance each earlier snapshot projected for it and how far that was from the actual balance.
Comparing forecasting methods for each category against its history is a different exercise, done when choosing how to fill out the horizon. That comparison tells you which method is most accurate; this page measures how far reality landed from what the company planned.
Frequently asked questions
How soon is there data?
From the first snapshot. The measurement matures over time: until there's a year of snapshots, the longest horizons have few observations, and the screen says so.
Can accuracy be measured retroactively?
Yes, if you have the estimates from past months. They're loaded and frozen as snapshots for those months, and then the current estimate is loaded to bring the forecast up to date.
Does measuring accuracy change the forecast?
No. This screen only reads. Estimates are still changed by their owners.
What counts as good accuracy?
It depends on the business. Payroll can be estimated more precisely than sales, so the company sets its own tolerances for receipts, payments and balances.