Treasury Cash-Position Forecasting at a US Regional Bank — 12-Basis-Point NIM Uplift
Treasury Predictor + Real-Time Visualizer giving the bank's ALM team a continuous, granular view of cash position that the previous month-end spreadsheets could not.
The challenge
The bank — a US regional bank with approximately $80bn in assets, retail and commercial deposit franchises and a meaningful treasury-services business — was running its asset-liability management function on a forecasting model that was structurally inadequate. The bank's cash-position forecast was a month-end spreadsheet exercise, refreshed weekly during normal periods and ad-hoc during stress periods, with a 30-day horizon at portfolio-aggregate granularity.
The model's accuracy was poor. Mean absolute percentage error against actuals was approximately 6.5% at the 30-day horizon, and worse at shorter horizons where the spreadsheet model had no behavioural signal for the specific deposit-customer cohorts driving the variance. The bank's treasurer was systematically over-provisioning liquidity buffer as a result, and the bank's CFO had calculated that the resulting opportunity cost was material to net interest margin.
The 2023 regional banking stress had concentrated the bank's focus on this problem. The bank had not been a stress casualty itself but had observed how quickly deposit composition could shift, and its CRO and CFO had concluded that the bank needed a fundamentally better understanding of its deposit-flow dynamics — not just for ongoing liquidity management but for the regulatory stress-testing the bank was now expected to perform with greater frequency and granularity.
The approach
MindMap deployed Treasury Predictor (Tp) as the cash-position forecasting engine, Real-Time Visualizer (Rv) as the treasurer's dashboard, Anomaly Detector (Ad) for early-warning detection on customer-segment deposit-flow shifts, and Data Lake Architect (Dl) for the underlying data infrastructure rebuild. The accelerator stack delivered an intraday 90-day cash-position forecast at customer-segment granularity, with continuous refresh and integrated stress-scenario simulation.
Phase one was the data-infrastructure rebuild. The previous spreadsheet model had been fed by manual data extracts; we built a streaming data ingestion from the bank's core deposit system, commercial-banking system, treasury-services platform, and the bank's Fed wholesale-funding portal. The ingestion populates a unified deposit-and-funding event store with full historical depth for behavioural modelling.
Phase two was the forecasting-model build. The model is an ensemble: a customer-cohort behavioural model (predicting deposit flows by customer segment, deposit type, balance band and tenure) using gradient-boosted trees on classical features, a macro-overlay model adjusting for rate-environment changes and competitive-rate dynamics using a deep-learning Temporal Fusion Transformer, and a wholesale-funding model integrating the bank's Fed funding capacity and counterparty limits.
Phase three was the treasurer-workflow integration. The treasury team's morning workflow now starts with the Real-Time Visualizer's overnight cash-position update and the day's recommended-action briefing. Stress scenarios — the standard regulatory scenarios plus bank-specific scenarios the treasurer configures — can be simulated through the model and the cash-position implications visualised.
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
Treasury Predictor
Customer-cohort deposit-flow and cash-position forecasting
Real-Time Visualizer
Treasurer dashboard with intraday refresh
Anomaly Detector
Early-warning detection on cohort-level deposit-flow shifts
Data Lake Architect
Streaming data infrastructure rebuild on Snowflake + Kafka
The architecture
The platform runs on the bank's AWS environment in the US East region. The deposit and funding data — meaningful PII exposure given the customer-segment granularity — stays inside the bank's VPC, with no data egress to external services.
The streaming ingestion uses Kafka as the event bus, with Debezium CDC connectors against the source systems. The event store is held in a combination of Snowflake (for the historical depth and the modelling workload) and a Redis-based hot tier (for the intraday model serving). The data layer handles approximately 8 million deposit and funding events per day at peak.
The forecasting model runs as a scheduled pipeline (model refresh every 4 hours during business hours, with the 90-day forecast rebuilt on each refresh) plus an event-triggered re-forecast for material deposit-flow events that breach defined behavioural thresholds. Model serving is on the bank's Snowflake Snowpark stack for the batch component and on a small Kubernetes deployment for the event-triggered path.
The customer-cohort segmentation is the architectural detail that produces the forecast accuracy gain. The previous spreadsheet model treated deposits as homogeneous within broad product categories; the new model segments deposits across approximately 240 distinct customer cohorts (combinations of customer segment, deposit product, balance band, tenure, rate sensitivity, primary-account indicator) and forecasts behaviour at the cohort level before aggregating.
The stress-scenario engine supports both deterministic stress scenarios (treasurer-configured rate shocks, deposit-runoff scenarios, wholesale-funding-disruption scenarios) and Monte Carlo simulation across the model's confidence distributions. The scenario engine outputs are persisted for the bank's regulatory stress-testing evidence base.
The numbers behind the story
Forecast accuracy has improved by approximately 5x against the previous spreadsheet model, measured as MAPE on the 30-day-horizon cash-position forecast. The 90-day horizon, which the previous model did not credibly support at all, is now produced with confidence intervals that the treasurer's team operates against daily.
The bank's liquidity-buffer carry has been reduced by approximately 8.4% in dollar terms, freeing balance-sheet capacity that has been redirected to higher-yielding assets. The CFO's measurement of the NIM impact is approximately 12 basis points annualised on the relevant balance-sheet base — a material contribution to the bank's overall NIM trajectory.
The bank's regulatory stress-testing process has been transformed. The previous quarterly stress-testing cycle absorbed roughly six weeks of treasury-team time; the new platform produces the standard regulatory stress outputs in days and supports the regulator's increasing expectations on stress-test frequency and granularity.
An unexpected outcome: the customer-cohort segmentation has identified deposit-stickiness patterns the bank's commercial-banking team had not previously had visibility into. The bank has used these insights to refine its deposit-pricing strategy at the customer-segment level, with measurable improvements in deposit retention on the segments the bank most wanted to retain.
“Our previous treasury forecasting was a weekly spreadsheet exercise with a thirty-day horizon and accuracy that frankly was not good enough for the rate environment we are now in. MindMap delivered a ninety-day intraday model with customer-cohort granularity, a measurable NIM uplift, and a stress-testing capability we were going to need anyway. The platform paid for itself in under a year.”— Treasurer· US Regional Bank
Why MindMap was chosen
The bank had evaluated two global treasury-analytics platforms and concluded that neither would deliver the customer-cohort-granular behavioural modelling that the bank's CFO believed was the source of the opportunity. The platforms were strong on the aggregate cash-flow forecasting but did not have the modelling depth at customer-segment granularity.
MindMap's Treasury Predictor accelerator had been deployed at a peer regional bank with a comparable scope and could demonstrate the customer-cohort modelling approach in production. The willingness to deploy entirely inside the bank's AWS VPC — including the model training, which involved customer-level deposit data — was a unique commercial position relative to the SaaS-only platforms.
The embedded treasury-domain expertise on the delivery team (two former bank treasurers and a former Fed examiner) was the third factor. The bank's CFO felt that the modelling team understood the regulatory and operational realities of bank treasury management, not just the modelling.
Related deployments
Sovereign WhatsApp Banking
ChatNext-powered WhatsApp bot deployed inside the bank's air-gapped data centre, handling balance, transfers, statements and loan applications in English and Swahili.
Cheque OCR at 99.1% Accuracy
DocuMage replaced a legacy template OCR for cheque clearing, processing 10,000 cheques per day at 99.1% field accuracy with 94% straight-through processing.
UK Challenger Bank KYC
OnboardX rebuilt the KYC pipeline with liveness, sanctions and PEP enrichment in a single STP flow, collapsing onboarding from 5 days to 4 hours.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.