Wealth DeskplusEnterprise Wealth Tech
Lending & Payments

Why Your Cash Forecast Still Misses: Seven Software Buying Mistakes and Their Fixes

A new platform rarely fixes a cash forecast that keeps missing. Here are seven buying mistakes that keep accuracy low after go-live, each with a fix and a question to ask vendors before you sign.

By the Wealth Deskplus editorial team · November 2026 · 9 min read

Forecast accuracy depends on inputs and ownership more than the engine

Say your team bought cash forecasting software in the last two years and the 13-week forecast still misses by double digits. The platform probably isn't the main problem. The usual causes sit before and after it. Bank data is incomplete. ERP feeds carry balances but not timing. Nobody keeps a record of how past forecasts performed. And nobody at the entity level owns the number.

This matters for your next decision. Teams that blame the tool often replace it, carry the same gaps into the new project, and pay a second implementation fee for the same result. Implementation is typically a meaningful fraction of the first-year subscription, sometimes more than the subscription itself.

Advertisement

Below are the seven mistakes that most often keep cash flow forecasting for treasury inaccurate after go-live. Each comes with a fix you can make on your current platform and a question to put to any vendor you're evaluating. They fall into four groups: data feeds (one and two), method and measurement (three and four), ownership and horizon (five and six), and how the deal is bought (seven).

Mistakes one and two: trusting the bank list and the ERP connector

Mistake 1: Assuming bank connectivity is solved because your banks are on the vendor's list

A bank logo on a "supported banks" page usually means at least one client has connected to that bank through some channel. It doesn't promise intraday data, every account, or every country. The channels also differ by region:

  • EBICS is common in Germany, France and Switzerland.
  • SWIFT, direct or through a service bureau, covers most multinational banking groups.
  • Host-to-host SFTP is common with banks in India, the Gulf and Southeast Asia.
  • Bank APIs work well at some large global banks and are patchy elsewhere.
  • Some smaller local banks in Africa and Latin America offer only a portal download.

Suppose a meaningful share of your cash sits in accounts that arrive by manual upload two days late. Your opening position is wrong before the forecast even runs.

Fix: Before you sign, build a connectivity map at the account level. For every account, record the bank, country and channel, whether you get prior-day or intraday data, and the file format (MT940, camt.053, BAI2 or CSV). Mark every gap.

Vendor question: "For our top 20 accounts by balance, which channel will you use, who builds the connection, what does it cost per bank, and how long did your last three similar connections take?"

Mistake 2: Letting the ERP feed carry balances but not dates

Many ERP connectors pull open receivables and payables totals. A forecast needs expected cash dates: invoice due dates adjusted for how each customer actually pays, payment-run calendars, approved purchase orders, payroll and tax dates. A connector that syncs nightly totals from SAP, Oracle, NetSuite or Microsoft Dynamics gives you a tidy dashboard and a weak forecast. Groups running several ERPs after acquisitions have it worse: half the entities feed in automatically and the rest send spreadsheets.

Fix: List the line-level fields you need: document date, due date, counterparty, currency, entity, payment terms and status. Also ask for 12–24 months of paid items, so the tool can learn how many days each customer actually takes to pay.

Vendor question: "Show us the field mapping for our ERP version, including how you work out expected payment dates for receivables. What happens with entities that run a different ERP?"

Mistakes three and four: one method for every flow, and no variance tracking

Mistake 3: Using one forecasting method for every cash flow

Vendors such as Kyriba, HighRadius and Trovata promote machine-learning forecasts. These models suit high-volume, repeating flows such as customer receipts spread across thousands of invoices. They do poorly with large known events like a dividend, a debt repayment or a tax installment. Schedule those; don't predict them. The reverse mistake is just as common: copying last month's receipts forward by hand ignores seasonal patterns your own history already shows.

Flow categoryExamplesBest methodOwner
Known or contractualDebt service, leases, dividends, tax installmentsScheduled from source contractsGroup treasury
High-volume and recurringCustomer receipts, routine supplier runsStatistical or machine-learning modelsSystem baseline, reviewed by the entity
Driven by business decisionsCapex milestones, large orders, bonusesSubmitted by the entityLocal finance
Rare or uncertainLegal settlements, M&A, asset salesManual entry with written commentaryTreasury plus deal or legal lead

Fix: Sort your flows into categories like these and give each its own method and owner.

Vendor question: "Can we set a different method for each category and entity, and override a model output with a recorded reason?"

Mistake 4: Running forecasts with no variance analysis

This one does the most damage because it hides the other six. If the system overwrites last week's forecast with this week's, you can't measure accuracy. Variance analysis needs frozen snapshots: one saved version per cycle, compared with actuals by category and entity for each horizon bucket (week one, weeks two to four, weeks five to 13).

Here's a hypothetical example. A group forecasts a net inflow of $4.0M for week six, and the actual figure is $2.6M. The miss is $1.4M, or 35%. Breaking it down shows $1.1M came from one entity, where two large customers paid late, and $0.3M came from the timing of a supplier run. The fix is to change the payment-timing assumption for those two customers. The model doesn't need changing.

Fix: Freeze a snapshot every cycle. Track accuracy by category, entity and horizon. Set tolerance bands, for example ±5% for week one and ±15% for weeks five to 13; these are illustrative, so set your own. Require written commentary whenever a variance breaks its band.

Vendor question: "How do you store forecast versions, and can we report variance by entity, category and horizon bucket inside the product, without exporting to Excel?"

Mistakes five and six: no owner at entity level, and one horizon for every decision

Mistake 5: Centralizing the forecast with no entity-level owners

In many groups, central treasury builds the forecast for 30 entities from system data, and local finance doesn't see it until a variance turns up. Yet local controllers know about the customer disputing an invoice, the tax audit settlement, the slowdown around Ramadan or Lunar New Year, or the slow government payment cycle. Without named owners, nobody stands behind the number.

Fix: Name an owner for each entity, and for each major category in large entities. Give them a submission workflow with deadlines and a review step, and show them their own variance results. Keep the task light: owners confirm or override the system baseline. They don't rebuild it.

Vendor question: "What does the entity submission workflow look like, how do you license occasional contributors, and can local users work in their own language and currency?" Watch the licensing answer. Per-user fees push companies to keep contributors out, and that brings the ownership gap straight back.

Mistake 6: Using one horizon for liquidity, funding and FX decisions

Different decisions need different forecasts:

  • Daily liquidity positioning covers the next one to ten days, account by account.
  • The 13-week cash forecast runs weekly by entity.
  • The 12–18 month funding and covenant view runs monthly and comes mostly from the budget (the indirect method).

Tools that are strong on direct, transaction-level forecasting can be weak on the indirect method. Stretching a 13-week direct model out to 12 months gives a precise-looking number that isn't reliable.

Fix: For each horizon, write down the decision it supports (cash sweeps, credit-line drawdowns, FX hedges, covenant tests), plus its granularity and method. Then reconcile where the horizons meet and against the FP&A plan.

Vendor question: "Which horizons do your current clients actually run in the product, and how does the long-range view reconcile to our budget?"

Mistake seven: buying on the demo, and what to test instead

Mistake 7: Signing after a demo on vendor data, with a price that hides connectivity costs

Every demo looks good on clean sample data. The quote can mislead too. Subscriptions usually scale with entities, bank accounts, users or modules, while bank connections, SWIFT service bureau fees and implementation partners often sit outside it.

Cost itemIndicative 2026 range (USD)Main cost drivers
Standalone or mid-market forecasting tool$20k–$100k per yearEntities, bank accounts, users
Enterprise treasury management system (TMS) with forecasting$100k–$400k+ per yearModules, entities, payment volumes
ImplementationRoughly 50–150% of first-year subscriptionNumber of ERPs, banks and entities
Bank connectivityA few hundred to several thousand dollars per bank per year; SWIFT service bureau fees often in the tens of thousands per yearChannel type, number of banks and countries

These are typical 2026 market ranges. They vary by contract, region and negotiating position, so confirm current pricing with each vendor.

Also check where the data will be hosted. The EU's GDPR, India's rules on storing payment data locally, and data rules in Saudi Arabia and the UAE can each limit which hosting regions you can use. Have counsel confirm what applies to your entities.

Fix: Insist on a pilot with your own data. Cover two or three entities, three months of history, and at least one bank for each channel type. Then run four to six weekly cycles with variance snapshots.

Vendor question: "Will you run a paid pilot on our data with our banks, and give us a fixed-fee quote for connecting all our accounts?"

A decision rule before you replace anything

If you already own a platform, freeze weekly snapshots and track variance for eight weeks before deciding anything. If the misses cluster in particular entities or categories where data is missing, fix the inputs and the ownership. Replace the tool only if it can't store forecast versions, can't apply different methods to different flows, or can't connect to the banks that hold most of your cash.

Quick answers

How accurate should our 13-week cash forecast be?

There's no universal benchmark. Accuracy depends on your industry, customer concentration and how much of your cash moves on contract terms. The most useful benchmark is your own history: track variance by horizon bucket for a quarter, then set tighter tolerances for the near weeks than the outer ones. A downward trend in variance tells you more than any single accuracy figure.

Do we need a full treasury management system, or is standalone cash forecasting software enough?

If your main gaps are visibility and forecasting across a modest number of banks and entities, a standalone tool is usually faster to implement and cheaper. A full treasury management system (TMS) makes more sense when you also need payments, in-house banking, debt and FX management, or hedge accounting in one controlled environment. Either way, check bank connectivity and variance tracking first, because they decide accuracy more than the size of the suite.

How long does implementation usually take?

Standalone forecasting tools often go live for the first entities within a few weeks to a few months. Full TMS projects usually take several months and sometimes longer. Bank connectivity and ERP mapping typically drive the timeline more than software setup, especially where banks need EBICS, SWIFT or host-to-host onboarding. Ask vendors for timelines from recent clients with a similar mix of banks and countries.

Are machine-learning forecasts worth paying extra for?

They can be, for high-volume recurring flows such as customer receipts, where there's enough history for the model to learn payment patterns. They add little for scheduled items like debt service or for rare events, and they can't make up for missing bank or ERP data. Test the model in a pilot on your own data and compare its variance with your current method before paying a premium.

Written by the Wealth Deskplus editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.