Skip to content

AI forecasting finance teams can explain, tested against last month's actuals.

We read the history and drivers a forecast learns from, test the first base case against last month's actuals, and a named person signs it.

Read about AI advisoryStart a conversation

AI forecasting for finance reads history and drivers, and the actuals decide what it can learn.

AI forecasting in finance means a model reading a group's monthly actuals and the drivers behind them to produce the next base case. Constancia, an EPM consultancy, builds it for multi-entity groups whose actuals reconcile to their ledgers each month.

Constancia's consultants have loaded actuals from entity ledgers on 58 implementations over 35 years, and a forecast reads the same load the close does. A forecast that reads a different load from the close explains a variance the close never saw.

Where 2 of 9 entities restate last month after the forecast has run, the board reads a forecast built on numbers the close has replaced.

A forecast passes through 5 steps each month, and the last one carries a signature.

The steps run in this order every month, and step 4 runs before anyone outside finance sees a number.

  1. 01

    The history it reads

    The model reads each entity's actuals by account and month, as the consolidation loaded them, and skips any month that failed to reconcile.

  2. 02

    The drivers

    Headcount, pipeline, volume and price each come from the system that owns them, with a named person outside finance behind each.

  3. 03

    The base case

    The model produces 1 base case by entity and account, and every figure in it traces to the history and the drivers it used.

  4. 04

    The test against last month

    Before the board sees it, the forecast is rerun as of 1 month ago and compared with the actuals that month produced.

  5. 05

    The sign-off

    The finance director signs the base case, and the model keeps the history, drivers and version it used under that signature.

A month that fails step 4 by more than the agreed tolerance goes back to step 1 with the reason recorded.

Reconciled actuals and unreconciled actuals give the same model 2 different forecasts.

The model is the same in both cases, and the actuals it reads make the difference.

A forecast on reconciled actuals

1 base case, and the finance director can say which month and which driver moved it.

Every month in the history agreed to its ledger before the model read it, and each driver has an owner who will explain a change. A question from the board about the revenue line ends at a named month and a named driver.

  • Actuals that agree to the consolidation, month by month
  • Drivers with owners in sales, HR and operations
  • A back-test against last month, kept with the forecast
A forecast on actuals that do not tie back

3 explanations for 1 variance, and the model is confident about all of them.

2 entities restated after the load, 1 entity's actuals arrive by email, and the model learns a history the close no longer recognises. A question from the board ends in a workbook that somebody has to open.

  • A history with 2 versions of the same month
  • A driver typed into a cell on the last working day
  • A back-test that fails for a reason nobody can name

A forecast built this way suits groups in which 4 conditions already hold.

  1. 01

    Actuals that reconcile every month

    Each entity's actuals agree to its ledger and to the consolidation before the month is closed.

  2. 02

    Drivers with owners outside finance

    A sales director stands behind the pipeline and an HR director behind the headcount when the forecast is challenged.

  3. 03

    A date the forecast has to be ready for

    A budget round, a funding round or a board meeting already fixed in the diary.

  4. 04

    A platform to run it on

    OneStream (EPM software), Abacum (FP&A software) or a decision to leave the spreadsheet estate by a date.

A single-entity business forecasting in 1 workbook has a working tool already, and this work waits until the entities multiply.

A group picks OneStream or Abacum according to where its actuals close.

Both platforms run the 5 steps, and the platform the actuals already close on is the one the forecast reads.

OneStream (EPM software)

SensibleAI on the consolidated actuals

  • For a group whose actuals close on the same platform the forecast reads
  • The history is the consolidation's own, so a back-test needs no second load
  • 1 entity list and 1 chart of accounts for the plan and the actuals
Abacum (FP&A software)

Abacum Intelligence on the FP&A model

  • For a finance team whose consolidation is settled elsewhere and loads in monthly
  • Drivers, base case and the back-test in the workspace the team forecasts in
  • Handed back to finance once the first signed forecast has run

We are an official partner of OneStream and of Abacum, with certified consultants on each.

The forecast shares its model with AI for FP&A and its checks with the agents.

See more on our

Subscribe on YouTube

  • Latest episode

    How an EPM implementation works, and why the software comes last

Talk to usSubscribe on YouTube

Every Sunday, 6am

The Sunday Close.

Our weekly newsletter on AI in finance.
Yours to enjoy with your Sunday morning coffee.

  • AI in finance observations from Alex (CEO) and Brad (COO)
  • Newest video from our YouTube channel
  • AI in finance trends we have spotted each week

Tell us the date you are working to.

The people who would do the work read every message and reply inside 2 working days.

Start a conversationRead about the readiness assessment