Cash Forecasting - Fundamentals¶

What It Is (General)¶

Cash forecasting is the practice of estimating a company's future cash position by projecting inflows (collections, financing, asset sales) and outflows (payroll, suppliers, debt service, capex). Forecasts can be short-term (daily/weekly for operational liquidity), medium-term (monthly for working capital planning), or long-term (quarterly/annual for strategic planning).

What It Means for Our ICP¶

How Treasury Teams Think About It¶

Treasury teams see cash forecasting as the foundation for all liquidity decisions. They distinguish between: - Direct forecasting: Transaction-based, built from known AR/AP, recurring payments - Indirect forecasting: Derived from P&L budgets, adjusting for non-cash items - Bottom-up (daily): For operational funding decisions - transactional level - Top-down (monthly): From FP&A for longer-term planning

The challenge is often reconciling these approaches and explaining variances.

Personio's Funding Challenge (Source: 2024-10-03) - Currently holding buffers on operational accounts because forecast isn't reliable enough - Need daily transactional forecast for just-in-time funding - Would prefer to deploy excess cash into investments rather than holding buffers

Sonder's Forecasting Process (Source: 2024-10-03) - Weekly forecasts, quarterly version saves - Don't overwrite forecasts with actuals - keep them separate - Compare accuracy over time: "How accurate were we at start of month vs mid-month?" - Kyriba for actuals, Google Sheets/Excel for forecasting - "Every forecast I build, I assume I'm gonna be wrong"

Sonder's Forecasting Philosophy (Source: 2024-11-19) - "Forecast is an opinion" - Treasury must apply judgment to inputs from other teams - 13-week rolling daily forecast, updated in ~15 minutes daily - Conservative on inflows: "Rather miss low than high" - upside surprises are better than shortfalls - Different touchpoints per team: calls with real estate, Google Sheets with tax, automated reports from payroll - Track accuracy over time: Snapshot comparison - "what did I think on Nov 1st about Nov 20th?" - High transaction volume = lower volatility - Many small transactions (hospitality) easier to forecast than lumpy B2B - Color coding convention: Black = actuals, Blue = manual inputs, Green = system-generated

Personio's Forecasting Philosophy (Source: 2024-11-26) - "Four core data points" for any forecast: value date, currency, account, amount - "without them it doesn't really help" - Categorization is bonus but valuable: Knowing if it's a lease payment vs general AP helps with variance investigation - Confirmed vs estimated framework: - Payroll, tax = confirmed (high certainty) - Customer inflows = estimated (lower certainty) - Approved invoices = confirmed; pending approval = estimated - Value date vs accounting date: Treasury cares about value date (when it actually debits); accounting may use invoice date or credit terms - Machine learning sweet spot: Outside business control (tax direct debits with date windows) vs business-driven (late invoice approvals) - Feedback loop philosophy: Granularity enables accountability; can show business the cost of late approvals - Budget vs transactional forecasting: Management cares about budget level; Treasury needs transactional level for efficiency

"Cash forecasting from a Treasury perspective is purely around efficiency. Having full visibility of flows in and out at a transactional level allows you to be fully efficient with your cash - that is then a value add for the business." - Tom Thorn

Typical Processes & Timing¶

Budget/forecast cycles: Many companies only do 1-2 formal forecasts per year
Rolling forecasts: More mature teams do monthly or even weekly updates
Entity consolidation: Decentralized groups struggle to aggregate entity-level data

Euroports' Structure (Source: 2025-10-27) - Multiple entities: Belgium (5), Spain (2), France - Want country-level consolidation - Currently limited to 1-2 budget/forecast cycles per year

Volvo Cars' Structure (Source: 2025-03-27) - ~40 entities report to Group Treasury - Business controllers submit monthly Excel forecasts by 3rd of month - Forecasts uploaded to Quantum (TMS) - SAP as common ERP across almost all entities - seen as critical - Day sweeps for cash pooling - entities don't hold cash - "I've seen big companies try to do forecasting with six or seven different ERPs and I don't think it really works so well" - Lee McEneff

Live Events/Festivals Group (Source: 2025-04-01) - 23 operating businesses (each may have multiple festivals) - PE-backed group; pressure to forecast cash for interest/debt planning - Unique cash cycle: Positive working capital - cash in (ticket sales) months before cash out (festival costs) - Inflows unpredictable (lineup quality, weather, ticket sales); outflows easier (stable costs) - Using external signals: Instagram engagement, post-event surveys, weather forecasts - Implementing TIS Payments + Integrity TMS for internal cash pooling - Bank rationalization: 40+ banks → ~20, local banking needed for physical cash deposits - "The revenue side... really needs that human touch to really get the vibe of how things are looking" - Dom Dubois

IHG (InterContinental Hotels Group) (Source: 2025-04-07) - ~6,500 hotels; only handful actually owned, most franchised or managed - Revenue = fairly predictable (royalties + management fees) - Always surplus USD (most revenue in dollars), need to fund subsidiaries in other currencies - ~10 Treasury accounts (Bank of America) in Kyriba; ~50 subsidiary accounts managed manually via India service center - Manual spreadsheet process: India sends spreadsheets, dealer collates into master file - Short-term focus: 6-12 weeks operational; long-term planning done at group level separately - "In 2025, this should be working better than it does... it should be easier, it should be more accurate." - Matthew Hook

Tools They Use Today¶

ERP systems - Source of AR/AP data, but often updated monthly (Mentioned by: Euroports)
Excel - Primary forecasting tool for most teams (Mentioned by: Euroports)
TMS platforms - Cash Analytics and others for visibility, limited forecasting (Mentioned by: Euroports)
Kyriba - TMS for cash forecasting; "nothing super smart, nothing of AI" - just open AR/AP + manual forecasts; no variance analysis capability (Mentioned by: ON)
HighRadius - AR/Collections management with customer payment pattern recognition; knows customer behaviors but doesn't flow to cash forecasting (Mentioned by: ON)
Dynamics - ERP with timing issues; paid items still show as open until accounting closes them (Mentioned by: ON)

ON's Kyriba Forecast Process (Source: 2025-11-17) 1. Start with initial balance (today's balance) 2. Add open AR (expected inflows) 3. Add open AP (expected outflows) 4. Add manual forecasts (salaries, taxes, significant payments) 5. Result = forecast for end of day/week/month "What Kyriba does is grabbing our initial balance, adding all of the expected outflows, expected inflows. Nothing super smart, nothing of AI."

ON's Forecast Source Preferences (Source: 2025-06-26) - Cash-ins: ML + maybe assumption for growth - Tax: Manual (from tax payment tracker spreadsheet) - Salaries: Manual - Fees (RCF, etc.): Manual - Open AR/AP: Combination (ML + ERP data) - Payment runs are weekly on Thursdays - even "ASAP" payments wait for next Thursday "In Kyriba or SAP, it's always a lot of work to update all the forecasts... you always have to take it back to Excel" - Amanda Mitt

ON's Data Architecture (Source: 2025-10-02) - Data flow: Dynamics 365 → BigQuery → Kyriba - BigQuery is cleaner data source than Kyriba (identified as preferred integration point for AP/AR) - Kyriba behavior: Deletes ALL forecast data daily and replaces - no smart updates, no historical snapshots - Budget codes (from Anaplan) mapped to Palm categories - Tax/talent files uploaded weekly - Palm's forecast handling: Only future dates included in forecast; past dates saved for variance analysis "The data goes to BigQuery... Palm could be here instead of Kyriba." - Amanda Mitt

Key Forecasting Principles¶

ON's Forecasting Philosophy (Source: 2024-11-19) - Forecast = starting balance + cash-ins - cash-outs = ending balance (fundamentally simple) - T+7 benchmark: Should be ~90% reliable based on booked data - Conservative approach: "Be pessimistic about cash-ins" - you have to pay on time but won't receive on time - Buffer calculation: Based on historical variance; more unforeseen events = bigger buffer - Cash pooling is a band-aid: "It's like giving candy to a kid every time it cries - you're not solving the root cause" - Lucía

Short-term vs Long-term: - Short-term = viability: Can company pay bills? Day-to-day operational focus - Long-term = performance: Investment allocation, strategic insights

What makes a forecast trustworthy: - ERP data as foundation (booked AP/AR) - Detailed bank account level info - Transparency on data sources (% from ERP vs manual vs ML) - Ability to explain model assumptions to stakeholders

"You have to be able to feel like you're contributing and understanding what the model is doing. Otherwise I wouldn't present this to my management." - Lucía (ON, 2024-11-19)

How They Talk About It¶

"Entity-by-entity forecasting" - forecasts at subsidiary level
"Group consolidation" - rolling up entity forecasts to corporate view
"Working capital movements" - AR/AP changes as forecast drivers
"Budget/forecast cycles" - formal planning periods
"Payment run" - scheduled batch payment processing (e.g., "every Monday for one AG")
"Cash application" - matching incoming payments to invoices (HighRadius term)
"Buffer logic" - if payment not made in last week, assume next payment run
"Open items" - AR/AP invoices not yet paid/received
"Model performance" / "Model explainability" - understanding and trusting forecast accuracy
"Save as" - creating quarterly forecast versions in Excel (Sonder)
"Seasonality" - year-over-year patterns; challenging for younger companies (Sonder)
"Zero variance = success" - "If you don't have to explain anything, it's so nice" (ON)
"T+7" - Seven day forecast horizon benchmark (ON)
"Forecast is an opinion" - Treasury must filter inputs through own judgment (Sonder)
"13-week rolling" - Standard forecast horizon for daily cash management (Sonder)
"Snapshot comparison" - Tracking historical forecast accuracy over time (Sonder)
"Confirmed vs estimated" - Tagging forecast certainty level based on data source (Personio)
"Value date" - When payment actually debits the account (vs accounting date or invoice date) (Personio)
"Four core data points" - Value date, currency, account, amount - minimum for useful forecast (Personio)
"Feedback loop" - Using granular variance data to hold business teams accountable (Personio)
"Rolled forward unpaid items" - When forecasted payment doesn't happen, roll it forward and flag it (Personio)
"Discrepancy reporting" - Report deviations/variances rather than overwriting figures (Volvo Cars)
"Pattern detection" - AI detecting when expected cyclical items are missing (Volvo Cars)
"Three wishes" - On-time reporting, discrepancy reporting, pattern detection (Volvo Cars)

Trust in AI for Forecasting¶

Volvo Cars on AI Trust (Source: 2025-03-27) "I wouldn't trust any less than a human... It's like a good Excel formula - once it works, like it works, you kind of start to trust a little bit more than yourself." - Lee McEneff

Payment Run Discipline¶

Levi's Process (Source: 2025-12-11) - Mid-week payment runs: Trigger payments Monday, pay Wednesday - allows time to fund and decide if postponing - Never pay on Fridays: No one does payment runs on Fridays - Calendar per input: Each forecast input (payroll, rent, AP by vendor type) has its own calendar/seasonality - Control what you can: "You can't influence when they're going to pay you, but you can certainly influence when you're going to pay out."

Top Customer Tracking¶

Levi's Approach (Source: 2025-12-11) - Focus ML forecasting on top 10 influential customers rather than every customer - "If you have an influential group, i.e. 10, then you go down to that level... you can actually apply machine learning for the forecast. This is how they pay." - Handle dilution (cancellations, shipping problems) separately from revenue

Bank Statements as ML Foundation¶

Levi's Insight (Source: 2025-12-11) "The bank statement is your key for your actuals and it's probably the best machine learning that you can actually use. It's a heavy lift. But if you have a system that could designate all your outflows and your inflows, half of your problem or challenges could be answered by that."

Industry Benchmarks & Market Reality¶

Treasury Dragons Audience Poll (Source: 2025-12-09) - 78% of treasurers say their medium-term forecast is "a good guide, but we could do better" - 8% are unhappy - "way out most of the time" - 4% are very happy - "usually very close" - 7% have no cash forecast at all

This validates that the vast majority of treasury teams see forecasting as an area for improvement.

Industry Expert View on AI & Forecasting (Source: 2025-12-09) "AI is there to help us as a tool. So we should not discount it, but do not trust it implicitly." - Royston Decoster (Ferguson, 37 years treasury experience)

"Shortened forecasts are becoming more reliable thanks to real-time feeds, but long-term forecasts still depend heavily on human judgment and scenario planning." - Royston Decoster

Confidence vs Accuracy¶

Palm Philosophy (Source: 2025-12-09, Treasury Dragons) "It doesn't matter how accurate your forecast is if you're not confident or trust in how it was generated, then it's a bit useless in a way, because you don't trust that data." - Gurjit Panu

This highlights that forecast value = accuracy + confidence. Users need to understand what's driving the forecast to trust it enough to act on it.

Blueprint for Liquidity Management¶

Industry Framing (Source: 2025-12-09, Treasury Dragons) "What people really want is a forecast that's more than just a report that's published out periodically, monthly. What they want is really a blueprint for how they manage liquidity. And they want to be able to do that confidently and they want to have the capabilities in there to be able to see around corners." - John Paquette (TIS)

Forecast Performance & Accuracy¶

The Trust Question¶

The central question is trust: "Can I trust it?" This comes from: - Internal stakeholders asking about forecast reliability - External peers (other treasuries) comparing notes - The need to make real decisions based on predictions

Treasury teams want a spectrum of detail: - Simple view: "Your forecast is good, you asked for 90% and we're hitting it" - Detailed view: "Which models are being used? What's the statistical error?"

"The answer also varies depending on who's asking. So to someone, you can just say, hey, it's good, we trust it, and they're fine. And some other people are like, no, but which models are you using?" - Lucia (ON)

Accuracy Measurement¶

How they measure: - Balance level (not cash flow level) - "The more is okay. Less is bad." - Target: 95% accuracy (ON internal Treasury goal) - Measured at quarter end for reporting

Key insight: Overestimating cash is acceptable; underestimating is risky (could lead to overdrafts)

"It doesn't matter if we forecast 25 million this week and we had 50 million... We need to forecast when it will happen, the expense, not a week after or a week before." - Rodrigo (ON)

Forecast Horizon Preferences¶

Horizon	Granularity	Why
1 week	Daily	Need to know exact timing for liquidity decisions
13 weeks	Weekly	Standard planning horizon; payment runs are predictable by day of week
3+ months	Monthly	"I don't need to know weekly in a year from now" - Yulia

Validation Methodologies¶

WMAPE (Weighted Mean Absolute Percentage Error) Method used by ON to track forecast accuracy: 1. Calculate MAPE per entity 2. Weight by entity's contribution to global cash position 3. Aggregate to get overall accuracy metric

ON's WMAPE Process (Source: 2026-01-22) - Weekly downloads from both Kyriba and Palm - Compare forecast vs actuals over 13-week horizon - Calculate weighted error based on entity contribution to global - Track week-over-week improvement in heat map format

Forecast Version Comparison - Compare different forecast versions for the same target week - Shows whether categorization changes improved accuracy - Heat map visualization: decreasing WMAPE = improving accuracy

Validation Timeframe - 6 weeks considered sufficient for initial validation (vs full 13-week) - Enough to get insights and build trust

"We validated over the span of six weeks. Even if it's not the 13 weeks, it's fine because it's already a good span to get some insights." - Federico (ON, 2026-01-22)

"Can I trust it?" - The fundamental question
"How freakish do you want to go" - Slider from simple to detailed metrics
"Feedback loop" - Using performance data to improve configuration
"Damify it" - Make complex model info digestible for non-technical users
"Model configuration" - Choosing which models/data to use per account
"WMAPE" - Weighted Mean Absolute Percentage Error (weights by entity's contribution to global cash)
"Forecast vs Forecast" - Comparing different forecast versions for the same target week

Sources: view all