Instacart Expert Feedback - 2025-07-01¶

TL;DR¶

Strong validation of scenarios/assumptions concept. Key insight: assumptions should be tracked over time so variance can be attributed to (1) actuals vs forecast and (2) assumption changes. Wants version saving, actuals overlay, and budget-to-cash translation support.

Overall Sentiment¶


Direction validated	Yes
Stage shown	Early prototype
Ready for next stage	Yes - build MVP
Blocker for them?	No (expert interview)

Source¶

Transcript: 2025-07-01-instacart-forecasting-discovery.md
Date: 2025-07-01
Participants: David Watt (Treasury, Instacart - formerly Sonder)
Context: Expert discovery interview - scenario/assumption studio prototype walkthrough

This Iteration: Validation¶

What they validated about the current design:

Apply assumptions to forecast - "Like a stair step. It would just go up and then stay flat again, right?" (payroll raise example)
Set effective date - "And when does it take effect?" → validated need for this
Visual preview of impact - "Charts are great because you can lay it out and say, well, here's what it's been, and then here's what I'm expecting"
Save as new baseline - "Apply it to your baseline forecast and now this is your new forecast"

Important Nuances¶

Assumptions need context/justification: "Based on updated July budget" with date and source
Visual sanity check is critical: "Wait a minute, why would it be... if it's been relatively flat, why would it now bounce up and down?"
Things move gradually in real life - "stair step" items are exceptions (new product/geography launch)
Bonuses are distinct from trends - Should be treated as explicit one-time events, NOT extrapolated from past data. "They might pay the bonuses in stock... or even if it is a cash bonus... typically paid in February or March for the prior year." This creates a disconnect: budget for this year includes bonus pool, but bonus paid is for last year.

Future Iterations: Suggestions¶

High Value¶

Version saving - "Save a version of the forecast and call this the July 1st forecast... so you can compare July forecast to June forecast"
Assumption change tracking - "The system's tracking all those things so it'd be pretty easy for the system to kick it back and say, well you added these, you changed these three assumptions over the course of the last period"
Variance attribution - Split variance into: (1) actuals vs forecast, (2) assumption changes. "That should give you the majority of your variance"
Actuals overlay - "I'd always want to see the last couple months of actuals if I'm going to change the forecast"
Manual override for one-offs - Explicit entry for bonus payouts, stock repurchases, vendor-specific payments that shouldn't be extrapolated from trends

Medium Value¶

Assumption context/notes - "Per David in team X on July 1st... so they could say well I didn't... sometimes you get into this like well we knew they were going to pay us like well you didn't tell me"
Budget comparison toggle - "Just pick which version of the forecast you want to compare to"
Unexplained variance bucket - After actuals + assumption changes, "there's always the unexplained variance"
Default smooth, allow step - "I found very few stair-step items in real life. Things move kind of gradually." UI should default to linear/smooth behavior but allow user-defined step changes for launches, bonuses, etc.

Longer Term¶

Budget-to-cash translation - Support translating P&L budget assumptions to cash forecast (revenue growth → collections with lag, bonus accrual → cash payment timing)
Scenario collaboration - "Collaborate with your colleagues on different assumptions"
Best case/worst case scenarios - "Save as base case scenarios... depending on your anticipated risk"

Questions Asked¶

When does assumption take effect? → Need effective date picker
Would actuals overlay help? → "For sure. I'd always want to see the last couple months"
Would you want assumption context saved? → Yes, for variance justification later

Raw Feedback Quotes¶

"Being able to save a version of the forecast and call this, this is the July 1st forecast, or the Q3 forecast or whatever you want to call it. So then you can compare the July forecast to the June forecast." - David Watt

"The system could easily tell you, right? Like, well, you added three assumptions over the course of June. And then your variance, you know, versus actuals... those would be your first two splits, right? How did I miss? Well, you changed your mind three times. And then also actuals came in different by this amount. And that should give you the majority of your variance." - David Watt

"The forecast is an opinion. Right about what the balance is gonna be. So I have to justify my opinions with facts, right? And well, here's why I had that opinion. It's very easy when you're doing it today, but... what was my opinion in March? That's very difficult to remember." - David Watt

"Charts are great because you can lay it out and say, well, here's what it's been, and then here's what I'm expecting... Wait a minute, why would it be... if it's been relatively flat, why would it now bounce up and down up and down? Like that doesn't make sense." - David Watt

"I find charts very helpful... Look at what the actuals have been recently and then see what... the assumption I'm changing and how it impacts things." - David Watt

"They might pay the bonuses in stock... or even if it is a cash bonus... typically paid in February or March for the prior year. So you'll have this disconnect... the budget for this year includes 10% for the bonus pool, but the bonus paid in this year is for last year." - David Watt

Key Use Case: Bonus Payments (One-Offs That Lag)¶

Bonuses are distinct from trends and should be treated as explicit one-time events, not extrapolated from past data: - Bonuses for prior year performance paid in Feb/March of current year - May be paid in stock (no cash impact) or cash - Budget accrual timing differs from cash payout timing - Requires manual entry/override, not ML prediction

Implication: Need to categorize these assumptions distinctly to support future reconciliation.

Key Use Case: Budget Reconciliation¶

David explained the challenge of aligning 13-week cash forecast with annual P&L budget:

Problem: - Budget is P&L/accrual basis (indirect) - Cash forecast is direct (collections/disbursements) - Need to translate budget assumptions to cash impact

Examples: - Revenue +10% → Collections +10% (but with payment term lag, e.g., 60 days) - Bonus accrual in budget → Cash payout Feb/March for prior year - Stock compensation in budget → No cash impact - New geography launch → Collections don't start until X weeks after launch

Quote: "That's the art of it is trying to... hey, here's our P&L budget for the year. And you have to say, well, what would that look like on a cash basis?"

Implication for Scenarios: Users may want to: 1. Import budget assumptions 2. Apply cash timing adjustments (lags, payment dates) 3. See how 13-week forecast aligns to budget year-end

Key Use Case: Payroll Annual Raise¶

Concrete example discussed: - Payroll historically: $1M/month - Annual raise: 3% starting end of July - Want: Apply "+3%" assumption starting July, see stair-step in chart - Quote: "If I know well there's everyone's getting a 3% raise at the end of July. Like how do I go in and bump it up by? Hey take June times 103%"

Action Items¶

[ ] Design version saving & comparison flow
[ ] Build assumption tracking for variance attribution
[ ] Add actuals overlay to scenario preview
[ ] Consider assumption notes/context field
[ ] Design one-off/manual override entry distinct from trend-based assumptions
[ ] Default to smooth forecast behavior, with explicit step-change option