What problem does it solve?
A hedging program is only as good as the exposure it hedges. When the forecast of future foreign currency flows is too low or too high, the company hedges the wrong amount and the gap shows up in the P&L as currency gains and losses it did not intend to carry.
The published cases describe the same starting point. ASML sells its systems in euros but pays some suppliers in US dollars, so its purchase hedging program rests on a forecast of expected USD material intake. According to the AFP case study, that forecast was only 70% accurate and came from a manual, Excel heavy process that was labor intensive and prone to error.
Microsoft Treasury, which hedges with FX forwards and options, used forecasts built from Excel spreadsheets fed by many groups across the company. Its Adam Smith Awards entry lists 80% of analysts' time spent on collecting and compiling data, and says the uncertainty in the forecasts led to variance between actual and forecast exposure and so to P&L volatility.
ASML says its model recognizes patterns and trends in historic actuals and uses them to predict future intake. The opportunity is to produce the exposure forecast automatically and more accurately, while hedge policy and execution stay with the treasury team. The ASML case dates from 2024; the Microsoft case is older, with sources from 2017 to 2019.
How does it work?
- Collect the actuals. Treasury and data science gather the history of the flows to be hedged, by currency, entity and period. ASML used five years of actual USD intake.
- Choose the model on held out data. ASML trained twenty open source Python algorithms on the first three years and tested them on the last two, kept the four best, retested them and fine tuned the winner. Microsoft built its accounts receivable forecast in Azure Machine Learning, which Treasury Today says is accessible by web for any user with access.
- Produce the exposure forecast without manual inputs. The model reads patterns and trends in the actuals and predicts future intake or collections per currency. EuroFinance's write up of Microsoft's award entry says its currency teams use machine learning that supplies forecasts in less than 15 seconds, where they used to be sourced from subsidiaries.
- Hedge against the forecast. Treasury applies its hedge policy to the forecast, for example layered forward cover, and places the trades with its banks or trading platform. The model does not decide ratios or execute.
- Compare with actuals every period. Each month the realised flows are added, forecast error is reviewed per currency, and the model is retrained. ASML says its model improves with each month.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Forecast accuracy | Too few to pool | 96% | 1 | 1 organization |
Value drivers: Risk and loss reduction, Employee productivity.
Indicative value
A manufacturer that buys USD denominated components and hedges EUR/USD exposure
USD 312,000 to USD 6.1 million
Expected cost of unintended currency exposure avoided, plus analyst time released per year
How this is calculated
Formula: annualUsd * errorReduction * mismatchCost + analystHours * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual USD denominated purchases hedged annualUsd, USD per year | 300,000,000 | 600,000,000 | Editorial assumption for the reference company, replace with your own hedged volume. |
| Reduction in forecast error, as a share of the hedged volume errorReduction, fraction | 0.05 | 0.2 | Editorial assumption, set well below ASML's reported move from 70% to 96% forecast accuracy, and not converted from that figure. |
| Average cost of an unintended open position, as a share of its value mismatchCost, fraction of the mismatched amount | 0.02 | 0.05 | Editorial assumption about the adverse currency move on the mismatch before it is corrected, replace with your own volatility view. |
| Analyst hours per year saved on collecting and compiling exposure data analystHours, hours per year | 200 | 600 | Editorial assumption, replace with your own time study. |
| Loaded cost of a treasury analyst hour hourlyCost, USD per hour | 60 | 100 | Editorial assumption, replace with your own loaded cost. |
What it leaves out: The first term is an expected value and a year with benign currency moves will show less. It leaves out hedging transaction costs, hedge accounting effort, the cost of building and running the model, and any change in the hedge ratio policy itself. The figures are not drawn from any one cited deployment.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
ASML
Netherlands · Manufacturing · 2024
ASML sells its chip making systems in euros but pays some suppliers in US dollars, so its purchase hedging program depends on a forecast of expected USD material intake. The treasury and data science teams replaced a manual, Excel based forecast with an automated machine learning model built on an open source Python algorithm, chosen after testing twenty candidates on five years of actuals. ASML reports forecast accuracy up from 70% to 96% and a monthly reduction in USD exposures.
- Forecast accuracy: 96%, forecast of USD denominated material intake
"With the AI model, forecast accuracy went up from 70% to 96% and USD exposures were reduced by $25-50 million monthly, making the hedging program significantly more effective."
Claimed by: organization
Microsoft
United States · Technology and software · 2017
Microsoft Treasury manages currency exposure across operations in more than 150 countries and hedges forecast exposures with FX forwards and options. Its forecasts used to come from Excel spreadsheets fed by many groups, and forecast variance fed P&L volatility. Treasury and the Cortana Intelligence team built a machine learning forecast for accounts receivable in Azure Machine Learning, and by 2019 currency teams used machine learning forecasts in balance sheet hedging. Treasury Today reports improved accuracy; EuroFinance reports a 25% lower currency impact on other income over 18 months, with no baseline given.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- At least three years of actual flows in each hedged currency, by entity and month
- A written hedge policy that defines the exposure, the horizon and the hedge ratios
- Known one off items (acquisitions, contract wins, price changes) tagged so they do not distort the pattern
Systems to integrate
- ERP purchase and sales order data, and bank statements for actuals
- Treasury management system or hedge register for existing positions
- Bank FX trading platform for execution, kept outside the model
Complexity: Medium
The build needs several years of clean actuals per currency from the ERP, a data science environment and agreement from treasury on what exposure is hedged and how far ahead. ASML built its model on free open source algorithms with its own data science team. Flows driven by one off events, such as acquisitions or large contracts, sit outside what history can predict.
- 1
Measure the current forecast
Record how accurate the manual forecast has been per currency over the last quarters and how many analyst hours it takes. ASML started from a 70% accuracy baseline, which gave it a clear target.
- 2
Test several algorithms on held out data
Train on the early years and test on the later ones, as ASML did with its twenty candidates. Compare the shortlisted models against the manual forecast on the same periods before choosing one.
- 3
Run in parallel before relying on it
Produce the model forecast next to the manual one for several cycles, review the gaps with the business, and agree with stakeholders before the hedge program switches over.
- 4
Wire the forecast into the hedge process
Feed the exposure number into the existing hedge workflow and keep trade approval and execution with treasury under the hedge policy.
- 5
Retrain and review every period
Add realised flows each month, track error per currency and set a threshold that triggers a model review.
Guardrails
- The model forecasts exposure; hedge ratios, instruments and trade execution stay with named treasury staff
- Forecast error per currency is reported next to the forecast, and a breach of the agreed band pauses reliance on the model
- Known one off events are entered by treasury as overrides, and each override is logged
- The model and its training data sit in the company's model inventory with an owner
KPIs to instrument
- Forecast accuracy per currency and horizon, against the manual forecast it replaced
- Unhedged or over hedged amount at period end, as a share of the realised exposure
- Analyst hours per month spent on collecting and preparing exposure data
- Currency impact on earnings versus budget
Human in the loop
Treasury reviews each period's forecast against actuals, adds one off events, and decides hedge ratios, instruments and timing. The model supplies the exposure number and no trade is placed from it without a treasury approval under the company's hedge policy.
Common failure modes
- History that no longer applies
- A change in suppliers, pricing or payment terms breaks the pattern the model learned, and the forecast looks precise while it drifts. Review error every month and retrain after structural changes.
- Accuracy measured on the wrong thing
- Reporting average accuracy hides large misses in the months that matter for the hedge. Track error per currency and per hedge date, and compare with the manual baseline.
- Forecast treated as a hedge instruction
- Staff hedge exactly to the number without a buffer for forecast error. Keep ratios and layers under the hedge policy, with cover below one hundred percent of the forecast.
What are the risks and rules?
EU AI Act
Minimal risk
Forecasting a company's own foreign currency flows is not listed in Annex III and makes no decision about a natural person. The model is used by treasury staff and does not interact with the public, so the obligations are limited to AI literacy (Article 4).
Rules that apply
Guidance
- NIST AI Risk Management Framework (NIST, North America). A structure to map, measure and manage forecast accuracy, drift and data quality risks of an internal forecasting model.
Controls to put in place
- Model inventory entry with an owner, the training period and validation results
- Hedge policy that limits the share of the forecast that may be hedged and sets approval levels
- Audit trail of forecasts, treasury overrides and approved trades
- Monthly comparison of forecast and realised flows reviewed by treasury and risk
Frequently asked questions
- How much more accurate can an AI exposure forecast be?
ASML reports that its USD material intake forecast went from 70% to 96% accuracy after it replaced an Excel process with a machine learning model, and says USD exposures fell by 25 to 50 million dollars a month. That is one company's own figure from its award entry. EuroFinance reports that Microsoft cut the currency impact on its other income by 25% over 18 months, with no baseline given.
- Does the AI place the hedges?
In the cited cases it forecasts the exposure that treasury hedges against. The sources describe no model that chooses ratios or executes trades, and the hedge policy and trade approval stay with treasury.
- Do you need a vendor or a data science team?
ASML says it built its model with its own data science team on free open source Python algorithms. Microsoft built its forecast with its Cortana Intelligence team in Azure Machine Learning. Both cited cases are in house builds, so these sources show no vendor involved.
How to cite this page
Blits.ai AI Use Case Library, "AI for FX exposure forecasting and hedging support", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/fx-exposure-forecasting-and-hedging-support. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published