Cash-flow & liquidity forecasting

Profitable companies still fail when they run out of cash, yet most liquidity forecasts are spreadsheets assembled weekly from bank portals, ERP exports and e-mails from subsidiaries. AI-based cash forecasting connects bank and ERP data automatically, categorises every transaction, learns payment behaviour and seasonal patterns, and produces rolling 13-week and 12-month forecasts with ranges rather than single numbers. Treasurers spend less time collecting data and more time acting on it (funding, investment, FX hedging, dunning). The same forecasting stack also serves the FP&A team's rolling P&L and budget forecasts.

Problem & value

• Who: CFOs, treasurers and controllers in mid-sized and large companies, owners and tax advisers of SMBs, public treasuries (municipal cash and short-term credit planning), and private households.
• Today: data is spread across ERP, banks, planning software and Excel. Forecasts rely on departmental knowledge, and accuracy is rarely measured against actuals afterwards. Microsoft names exactly these problems as the reason for its ML-based cash forecast in Dynamics 365 Finance [1].
• AI value: automated data collection and transaction categorisation, statistical/ML forecasts for the "uncertain" flows (customer payment timing, card and e-commerce receipts, variable costs), forecast-vs-actual tracking, and early warning when headroom falls below a threshold.
• Legal driver (DE): § 1 StaRUG obliges managing directors to continuously monitor developments that could threaten the company's existence [2]. Imminent illiquidity under § 18 InsO is judged against a forecast period that is "as a rule" 24 months [3]. A reliable, documented liquidity forecast is therefore a governance duty, not only a convenience.
• Reported results (vendor-reported, not independently verified): HighRadius claims up to 95 % forecast accuracy and quotes an unnamed treasury customer with 20 % better accuracy and daily effort cut from 2 hours to 15 minutes [4]. TIS quotes TomTom with "70 % less time on cash forecasting" [5]. Always validate such claims on your own history, because definitions of "accuracy" differ.

Variants

• SMB / Mittelstand: a cash-flow tool on top of bank feeds and accounting data, with a 13-week rolling forecast, scenarios and a tax-adviser view (Agicap, DATEV-based tools, banking apps).
• Enterprise treasury: multi-entity, multi-currency forecasting in a treasury management system (TMS) or in SAP Cash Management, with in-house bank, cash pooling and FX exposure feeds.
• FP&A rolling forecast (merged variant): revenue, cost and budget forecasting as a rolling P&L forecast with predictive planning, for example SAP Analytics Cloud Predictive Planning, Jedox, Anaplan PlanIQ or Board. It uses the same time-series stack and serves the same finance buyer. The planning-cube workflow, what-if modelling in plain language and variance commentary are covered in DATA-024 and DATA-002.
• Receivables-driven forecast: invoice-level payment-date predictions (PRED-021) are aggregated into the inflow line. This is usually the largest gain in accuracy for B2B companies.
• Banks & insurers: intraday and structural liquidity forecasting for liquidity-risk management. This is the same technique under stricter model-risk governance and DORA ICT rules.
• Public sector: municipal and state cash management (tax receipts, transfers, short-term credit). Macro tax-revenue forecasting is covered in PRED-047.
• Private use: "safe-to-spend" balance prediction and recurring-payment detection in banking and budgeting apps (PERS-001).

How it works

1. bank feeds (camt.053 / PSD2 API / EBICS) + ERP open items + plans
2. normalise & categorise transactions (rules + ML/LLM classifier per cash-flow category)
3. split flows: known/contracted (payroll, loans, tax, open AP/AR) vs. uncertain (new sales, card receipts)
4. known flows: deterministic schedule; AR: payment-date prediction per invoice (PRED-021)
5. uncertain flows: time-series models per category/entity/currency (stat + GBM + foundation model ensemble)
6. reconcile across hierarchy (category
7. entity
8. group), convert FX
9. prediction intervals, scenarios & stress tests
10. treasurer review & manual overrides (human-in-the-loop)
11. publish: dashboard, headroom/covenant alerts, LLM-written commentary
12. forecast vs. actual tracking
13. retraining

• Forecast engine: for daily/weekly cash categories, strong baselines are ETS/ARIMA/Theta, then gradient boosting with calendar features (month-end, holidays, payroll days, tax dates), and increasingly zero-shot time-series foundation models with covariates (Chronos-2, TimesFM) in an ensemble.
• Categorisation: classifiers trained on historic bookings, with LLM fallback for unusual remittance texts.
• LLM layer: answers questions such as "why is week 32 lower than last forecast?" from structured variance data, drafts the weekly liquidity memo and runs scenario requests via tools/MCP. The LLM never produces the numbers itself.
• Human-in-the-loop: treasurers approve overrides (e.g. a known large customer payment), and every funding, investment or hedging decision stays with people under the treasury policy.

Data preparation

Cash-flow forecasts need bookings and open items: receivables and payables from accounting or ERP (DATEV, SAP, Lexware), bank statements (EBICS, FinTS or PSD2 feeds), payroll and tax calendars, contracts and leasing schedules, and sales pipeline or order backlog. In small companies much of it is in Excel, on paper invoices or only in the tax adviser's system.

Typical gaps: bank transactions without clear invoice reference, unmatched payments, duplicate or missing open items, inconsistent customer names, free-text payment terms, recurring payments hidden in bulk postings, and no record of actual payment behaviour per customer.

Steps: connect bank feeds and export open items (PREP-005), digitise paper invoices (PREP-006) and extract their data (DOC-001), match payments to invoices to derive real days-to-pay per customer, unify customer names (PREP-015), categorise cash flows with a simple scheme (PREP-019), and keep a back-test set of past months (PREP-018). Tools: bank connection of the cash-flow tool, Power Query, pandas, n8n. See DATA-001, DATA-002 and GUIDE-003.

Effort: about a third of the project. Clean data from the last 12 to 24 months is usually enough; seasonality needs at least two years. Agree with the tax adviser on export interfaces.

Implementation paths

Which path for whom

• smb: A. Mature, EU-hosted cash-flow tools (for example Agicap) connect to bank and accounting and run standard processes at SMB prices.
• enterprise: B. Treasury is core, ERP and bank data are specific, and forecasts must feed an existing TMS, so integrate cloud ML or API forecasting with your own data.
• developer: C. Open forecasting libraries (StatsForecast, Prophet-style, foundation time-series models) run on HW1 with full control.
• public (municipalities, agencies): C. Budget and payment data are sensitive and procurement favours open stacks on sovereign cloud.
• private: A. Budget and finance apps cover household liquidity without own infrastructure.

Path A — Buy (off-the-shelf)

• SMB / mid-market: Agicap (EU, data hosted in Google Cloud Belgium; 13-week rolling forecasts, scenarios, Agicap AI, Claude connector) [6]; liquidity tools in accounting software and banking portals.
• Enterprise TMS with AI forecasting: Kyriba (Cash Management AI since 2022, newer agentic assistant TAI) [7][8], HighRadius (AutoML across 100+ models) [4], TIS (DE, Walldorf; AI Cash Forecasting, AI Assistant) [5], Coupa Treasury (formerly Bellin, DE), Serrala (DE, Hamburg; cash management and liquidity planning embedded in SAP) [9], Trovata (bank-API-native) [10], GTreasury.
• ERP-native: SAP S/4HANA Cash Management with AI-based liquidity planning [11]; Microsoft Dynamics 365 Finance insights cash-flow forecast (time series via AI Builder, snapshot what-ifs) [1].
• FP&A: SAP Analytics Cloud Predictive Planning, Jedox (DE; JedoxAI, Planning Wizards for revenue, cash-flow and rolling forecasts) [12], Anaplan PlanIQ, Board.
• Choose A when bank connectivity, ERP connectors and audit trails are the hard part, which is usually the case. Check EU hosting, EBICS/camt support, the DATEV/SAP connector and whether forecast accuracy is measured transparently.

Path B — Build on cloud APIs

• Forecasting services: Google BigQuery ML (TimesFM, incl. TimesFM 3.0 for commercial use) [13], Amazon SageMaker/AutoGluon-Cloud with Chronos-2 [14], Azure Machine Learning AutoML forecasting. Use EU regions (Frankfurt, Belgium, Netherlands, Sweden).
• LLM layer: an EU-hosted frontier model (Mistral, Claude, GPT or Gemini in EU regions) for commentary and Q&A over a semantic layer, with tool calls into the forecast store.
• Data: bank aggregation via PSD2 AISP providers (e.g. finAPI, Tink) or EBICS, ERP extraction, and a warehouse (BigQuery, Snowflake, Databricks).
• Choose B when you have an in-house data team and want forecasts inside your own BI.

Path C — Self-host (open source)

• Stack: Nixtla StatsForecast (AutoARIMA, AutoETS, Theta, MSTL; large batch speed) [15], MLForecast/LightGBM, AutoGluon-TimeSeries 1.6 (ensembles of stat, tree, DeepAR/TFT and Chronos models, probabilistic output) [16], Chronos-2 (Apache-2.0, 120M parameters, covariates, multivariate) [14][17]. TimesFM 2.5 weights are Apache-2.0, while TimesFM 3.0 weights are licensed for non-commercial use only [13].
• Orchestration and UI: Airflow/Dagster, PostgreSQL, Superset/Metabase dashboards. Use a small local LLM (e.g. Mistral Small, Qwen3 8–32B via vLLM/Ollama) only for commentary.
• Hardware: HW1. A CPU server handles thousands of cash series, and Chronos-2 runs on CPU. HW2 is needed only if the local LLM commentary layer runs on the same machine.
• Ops effort: moderate, because the bank/ERP connectors, not the models, cause most of the work.

Path D — Custom training

It is rarely worth it. Fine-tuning a time-series foundation model on your own long, high-frequency cash history (e.g. thousands of entity-account-category series over several years) can help large groups or banks with intraday liquidity. Most companies gain more from better data and from invoice-level AR prediction (PRED-021) than from custom deep models.

State of the art (2026)

• Time-series foundation models are mature. Chronos-2 (Oct 2025) supports past and known-future covariates and multivariate forecasting, and its authors report the best results among pretrained models on fev-bench and GIFT-Eval [17]. Google released TimesFM 3.0 in August 2026 with covariate support, but under a non-commercial weights licence (commercial use via BigQuery ML) [13].
• Hybrid "direct + indirect" method: contracted flows are scheduled deterministically, AR is predicted per invoice and residual categories are forecast statistically. This beats a pure top-down series model in practice.
• Agentic treasury: TMS vendors ship assistants that query cash positions, explain variances and draft reports (Kyriba TAI, TIS AI Assistant, Agicap's Claude connector, Jedox MCP) [5][6][8][12].
• Limits: cash flows are lumpy and event-driven (large orders, M&A, tax audits, customer insolvency). Models cannot foresee structural breaks, so wide intervals and scenarios are honest outputs. Data quality (uncategorised transactions, missing subsidiary data) is the main bottleneck. LLM explanations must be grounded in computed variances.

Shared resources & combinations

• Same time-series stack (time-series + tabular-ml, HW1): PRED-001 demand, PRED-002 sales forecast, PRED-008 price forecasting, PRED-011 workload. One forecasting service (StatsForecast/ AutoGluon/Chronos-2) can serve all of them.
• Upstream data: AGENT-023 (cash application gives clean AR), PRED-021 (payment-date prediction feeds inflows), DOC-001 (AP invoices give outflows), AGENT-006 (reconciled bank data).
• Downstream: DATA-024 FP&A planning and scenarios, DATA-002 report narratives, DATA-001 conversational BI on cash data, ANOM-006 ledger anomalies for unexpected outflows.
• Shared LLM/agent layer with other finance copilots. Shared workflow automation (n8n, Power Automate) for bank-file ingestion.

Effort, cost & KPIs

• Path A: SMB tools in days to weeks. An enterprise TMS often takes several months, mostly for bank connectivity and ERP mapping. Cost is driven by entities, bank accounts and modules.
• Path B/C: 2–4 months for a production pipeline with an internal data team. Compute cost is negligible, and engineering and data cleansing dominate.
• KPIs: forecast accuracy per horizon (WAPE/MAPE on weekly net cash flow, bias), interval coverage, hours spent per forecast cycle, share of auto-categorised transactions, idle cash and short-term borrowing cost, covenant breaches avoided, DSO (with PRED-021).

People & change

Start with management and the department lead (finance, controlling or treasury); they own forecast and decisions. Involve accounting staff as end users and data owners: they know unusual payments and customers. IT or the tax adviser sets up bank and ERP connections. Customers and suppliers are not involved, but payment terms may change as a result.

Works council: usually no. Forecasts are made on company-level cash flows, not employee behaviour. It becomes relevant if the tool logs user activity (§87(1) no. 6 BetrVG, as for most software) or if payroll data enters at employee level; inform the works council anyway under §90(1) no. 3 when work processes change.

Data protection: invoices and payments of sole traders and individuals are personal data; sign an AVV with the tool vendor and check bank-connection terms (GUIDE-005, GUIDE-010). Ask the DPO; a DPIA is usually not required. AI literacy: a short session on forecast ranges and scenario thinking supports Art. 4 duties (GUIDE-006).

Pilot with one entity or account group for three months, and compare forecast and actuals every week. Typical concerns are "the forecast is wrong" and loss of control; answer with transparent assumptions, editable scenarios and a human who signs off before any financing decision. See GUIDE-009 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: minimal risk. Corporate liquidity forecasting is not an Annex III use. Creditworthiness scoring of natural persons (Annex III 5(b)) is a different use case (PRED-009). Art. 4 AI literacy applies. Chatbot-style assistants that interact with people fall under Art. 50 transparency [18].
• GDPR: mostly company data, but bank transactions include payroll, sole traders and private customers. Pseudonymise counterparties and sign a DPA with the vendor. PSD2 account access requires a licensed AISP. Private budgeting apps process sensitive personal financial data.
• Corporate law: document the forecast method and assumptions (StaRUG § 1 monitoring, § 18 InsO 24-month horizon) [2][3]. Directors must be able to show how the AI forecast was reviewed.
• Financial sector: banks, insurers and payment institutions must treat TMS/forecast SaaS as ICT third-party providers under DORA, and liquidity-risk models fall under supervisory model governance.
• Risks: over-trust in point forecasts, silent data gaps, and vendor accuracy claims measured on other definitions. Mitigate with intervals, backtesting, override logs and a manual minimum-liquidity floor.

Getting started

1. Inventory cash data: bank accounts, connectivity (EBICS/API), ERP open items, and recurring obligations (payroll, tax, loans).
2. Define cash-flow categories and a 13-week rolling template. Start measuring forecast vs. actual now, so you have a baseline.
3. Pilot one Path A tool (Agicap for SMB, an SAP-integrated TMS for enterprise), or a StatsForecast/ Chronos-2 notebook on 2–3 years of categorised bank history.
4. Add invoice-level AR payment prediction (PRED-021) for the inflow side, plus scenarios and headroom alerts.
5. Document method, overrides and review in the treasury policy. Then extend to the 12/24-month horizon and the FP&A rolling forecast.

Sources

1. Microsoft Learn — Cash forecast (Finance insights, Dynamics 365 Finance) — https://learn.microsoft.com/en-us/dynamics365/finance/finance-insights/cash-flow-forecast-intro
2. § 1 StaRUG (crisis early detection) — https://www.gesetze-im-internet.de/starug/__1.html
3. § 18 InsO (imminent illiquidity, 24-month forecast period) — https://www.gesetze-im-internet.de/inso/__18.html
4. HighRadius — Cash forecasting software — https://www.highradius.com/product/cash-forecasting-software/
5. TIS (Treasury Intelligence Solutions) — AI Cash Forecasting — https://tispayments.com/
6. Agicap — cash flow management for SMBs and mid-market — https://agicap.com/en/
7. Treasury Management International — Kyriba launches AI-driven cash management (2022) — https://treasury-management.com/news/kyriba-launches-ai-driven-cash-management
8. Kognitos — Top AI cash-flow forecasting tools for treasury 2026 (industry overview, vendor claims) — https://www.kognitos.com/blog/top-ai-cash-flow-forecasting-tools-treasury-2026/
9. Serrala — finance automation & SAP-embedded cash management — https://www.serrala.com/
10. Trovata — bank connectivity and cash forecasting — https://www.trovata.io/
11. SAP — SAP S/4HANA Cloud for cash management — https://www.sap.com/uk/products/financial-management/cash-management.html
12. Jedox — AI planning platform — https://www.jedox.com/en/platform/ai/
13. Google Research — TimesFM repository (versions, licences, BigQuery ML) — https://github.com/google-research/timesfm
14. Hugging Face — amazon/chronos-2 model card — https://huggingface.co/amazon/chronos-2
15. Nixtla — StatsForecast documentation — https://nixtlaverse.nixtla.io/statsforecast/index.html
16. AutoGluon — Time series forecasting docs — https://auto.gluon.ai/stable/tutorials/timeseries/index.html
17. Amazon Science — chronos-forecasting repository (Chronos-2, benchmarks) — https://github.com/amazon-science/chronos-forecasting
18. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj