Cash Forecasting¶

Overview¶

Cash forecasting is the process of predicting future cash inflows and outflows to ensure a company has adequate liquidity to meet its obligations. For treasury teams, this ranges from short-term daily/weekly forecasts (operational) to longer-term monthly/quarterly projections (strategic planning).

Accurate cash forecasting enables treasury teams to optimize working capital, make informed investment decisions, manage debt facilities, and avoid costly surprises. Poor forecasting leads to either excess idle cash (opportunity cost) or dangerous liquidity shortfalls.

Key challenges include decentralized data sources, infrequent forecast updates, difficulty reconciling forecast vs actual, and limited visibility into working capital drivers. Palm can help by aggregating data across entities, enabling entity-level forecasting with group consolidation, and providing variance analysis tools.

For detailed ICP context and terminology, see fundamentals.md

Top Jobs & Desired Outcomes¶

Full history: jobs.md

1. Explain and demonstrate forecast reliability to stakeholders ✓¶

Desired Outcomes: - Minimize the effort required to demonstrate forecast accuracy - Increase the transparency of model selection and performance - Reduce skepticism from finance stakeholders about ML-based forecasts - Increase ability to answer "can I trust it?" with data

Sources: ON x4 (Confirmed) - "Can I trust it?" / "You have to feel like you're contributing and understanding what the model is doing"

2. Configure forecasting models based on data quality ✓¶

Desired Outcomes: - Minimize forecast errors caused by poor data sources - Increase ability to disable/enable ML per account based on suitability

Sources: ON x2 (Confirmed) - "For this account, please deactivate machine learning"

3. Produce a short-term cash forecast I can trust and act on ✓¶

Desired Outcomes: - Minimize the time spent updating daily forecast (target: 15 minutes or less) - Increase the percentage of forecast based on booked/actual data vs guesses - Reduce the frequency of surprise payments not captured in forecast

Sources: ON x2, Sonder, Levi's (Confirmed) - T+7 benchmark: ~90% reliability; "We try to do it in 15 minutes or less"

4. Optimize use of operational cash ✓¶

Desired Outcomes: - Minimize buffer balances held on operational accounts - Increase investment income by deploying excess cash - Reduce uncertainty about daily funding needs

Sources: Personio x2, Sonder (Confirmed) - "Cash forecasting from a Treasury perspective is purely around efficiency"

5. Forecast cash flows entity-by-entity across decentralized group structure ✓¶

Desired Outcomes: - Minimize the time spent chasing entities for forecast submissions - Reduce errors from manual Excel manipulation (format, decimal conventions) - Increase visibility into forecast changes vs original figures

Sources: Euroports, Volvo Cars (Confirmed) - 40 entities at Volvo; "When it's on the Excel, anything can go wrong"

6. Improve forecast accuracy by handling outliers ✓¶

Desired Outcomes: - Minimize the impact of one-off large payments on future forecasts - Increase confidence in forecasts by excluding known anomalies

Sources: ON, Sonder (Confirmed) - "Algorithms have a hard time with sudden changes"

Key Pain Points¶

Full history: pain-points.md

Can't easily explain forecast accuracy - Hard to answer "can I trust it?" from stakeholders (Source: ON)
No visibility into model selection - Don't know which models are being used or why (Source: ON)
Excel-based validation workflows unsustainable - Manual WMAPE tracking hard to maintain (Source: ON)
No built-in forecast version comparison - Have to manually compare different forecast versions (Source: ON)
150 accounts blindness - "You become fully blind. You don't know where to draw your attention." (Source: ON)
Holding buffers because forecast isn't reliable enough - Can't do just-in-time funding (Sources: Personio x2)
Manual spreadsheet process - No system, everything basic and manual (Source: Personio)
Last-minute surprise payments - Payments not in anyone's radar, discovered day-of (Source: ON)
No smart forecasting in TMS - Kyriba just does open items, no ML/patterns (Source: ON)
Outliers skew forecasts - Large one-off payments distort future predictions (Sources: ON, Sonder)
Decentralized entities make consolidated forecasting difficult - Multiple entities need individual and group views (Sources: Euroports, Volvo Cars)
Data fragmentation industry-wide - "The biggest pain is in gathering all the data into one platform" (Source: Treasury Dragons - Cobase)
Spreadsheet default despite tools - 78% of treasurers want better forecasts; many revert to Excel (Source: Treasury Dragons - Ferguson)

Key Opportunities¶

Forecast explainability/trust - Answer "can I trust it?" with model performance visibility
Outlier handling - Mark transactions as one-offs to exclude from learning
Entity-by-entity forecasting with group consolidation - Enable local input with central visibility
Category-level views - See development of salaries, Capex, etc. separately
AR pattern integration - Apply customer payment behaviors to collection forecasts
Pattern detection - Proactively detect missing cyclical items (e.g., UK tax every March)
Discrepancy reporting - Report variances vs overwriting forecasts to improve accountability
Top customer tracking - Focus ML forecasting on influential top 10 customers rather than all
Bank statement ML foundation - Use bank statements as primary ML data source
Push button forecasting - One-click weekly forecast generation
Confidence + accuracy - Users need to trust HOW forecast was generated, not just see the number
Blueprint for liquidity management - Forecast as dynamic decision tool, not static report

Open Questions¶

[ ] What's the right balance between local entity input and central forecasting?
[ ] How to handle factored receivables in cash forecasting?
[ ] How to surface model performance without overwhelming non-technical users?
[ ] How to implement "calendar per input" for different seasonality patterns?

Last updated: 2026-02-17 | Sources: 26 transcripts (view all)