The franc floor’s three calm years were the most dangerous data in the risk system

When Switzerland dropped its euro floor and the ECB launched bond buying, models built on a managed price failed in four ways: the history, the number range, the price path and the client. Each was a data assumption, not bad luck.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

When Switzerland dropped its euro floor and the ECB launched bond buying, models built on a managed price failed in four ways: the history, the number range, the price path and the client. Each was a data assumption, not bad luck.

  1. The week The week of 19 January 2015 opened in the wreckage of the previous Thursday.
  2. The history: three years of a managed price From September 2011 to January 2015, the euro-franc exchange rate could not fall below 1.20, because the SNB stood ready to buy unlimited foreign currency to defend it (SNB).
  3. The number range: code written for positive numbers The SNB’s decision pushed its deposit rate to minus 0.75 percent and its Libor target range as low as minus 1.25 percent (SNB).
  4. The path: prices that jump instead of move The third assumption is about how prices get from one level to another.
  5. The client: borrowed exposure that passes losses upstream The fourth assumption concerns who bears the loss.
  6. Where this framing is weakest The strongest objection is that the SNB’s move was a genuine surprise, and no data discipline could have predicted its timing.
  7. What data and risk leaders can build The four assumptions point to four data products, each with an owner and a measurable output.
  8. With hindsight Regulators moved first on the market risk rules.

The week

The week of 19 January 2015 opened in the wreckage of the previous Thursday. On 15 January the Swiss National Bank had announced that it was discontinuing the minimum exchange rate of 1.20 Swiss francs per euro, with immediate effect. At the same time it cut the interest rate on sight deposit balances above an exemption threshold by half a percentage point, to minus 0.75 percent. It moved its target range for three-month Swiss franc Libor further below zero, to between minus 1.25 and minus 0.25 percent (SNB).

The floor had been in place since September 2011. The SNB described it as an exceptional and temporary measure, introduced when the franc was massively overvalued and markets were highly uncertain. It said the overvaluation had eased since then. It pointed to growing divergence among major central banks and the euro’s fall against the dollar, and concluded that enforcing the floor was no longer justified. The rate cut was meant to stop the exit from tightening monetary conditions too much (SNB).

The market move was violent. Bloomberg reported that the franc surged as much as 41 percent against the euro on the day, its biggest gain on record. It rose more than 15 percent against every one of the more than 150 currencies it tracked (Bloomberg via swissinfo.ch).

The losses landed first on retail currency brokers and their clients. FXCM was a broker listed on the New York Stock Exchange. On 15 January it said that client losses had left negative equity balances owed to the firm of about $225 million, and that as a result it might be in breach of some regulatory capital requirements (FXCM). Bloomberg reported that Alpari UK said it had entered insolvency, citing exceptional volatility and an extreme lack of liquidity. Global Brokers NZ closed. Swissquote set aside 25 million francs for client negative balances, and IG Group estimated its impact at up to £30 million (Bloomberg via swissinfo.ch).

FXCM announced a $300 million financing from Leucadia National on Friday 16 January to restore capital at its regulated subsidiaries. When the terms became public early in the week of 19 January, the market judged them harsh and FXCM’s shares collapsed. Bloomberg also reported disruptions at larger institutions: Deutsche Bank’s Autobahn electronic platform briefly stopped quoting, and HSBC was looking into reports of customers buying francs at off-market rates (Bloomberg via swissinfo.ch).

Other pegs came under pressure the same week. Denmark’s central bank, which holds the krone in a narrow band against the euro, cut its certificate of deposit rate on 19 January. On 22 January it cut again, by 0.15 percentage points to minus 0.35 percent effective the next day, after buying foreign currency in the market (Danmarks Nationalbank).

Then, on Thursday 22 January, the European Central Bank announced an expanded asset purchase program. Combined monthly purchases would reach €60 billion, now including bonds issued by euro area central governments, agencies and European institutions, bought in the secondary market. The purchases were intended to run until at least September 2016, and in any case until the Governing Council saw a sustained adjustment in inflation toward its aim of below but close to 2 percent. Only 20 percent of the additional purchases would be subject to loss sharing among Eurosystem central banks (ECB).

Taken together, the two central bank decisions told risk teams that a pegged currency could float overnight and that negative yields would spread and persist. The ECB move had been widely expected for weeks. Almost nobody had anticipated the SNB’s. Both changed the data on which banks, brokers and asset managers ran their models.

The franc shock of January 2015 is usually told as a market story: a central bank surprised everyone and some brokers failed. We read it as a data story. The losses came from four assumptions that sat inside risk data and models, each of which had been true for years and stopped being true in a morning. None was a forecast. Each was a property of the data that nobody had labeled.

The four assumptions are distinct, and together they account for the ways models failed that week:

  • The history: the past three years of prices described the future, even though a central bank had been setting the price.
  • The number range: interest rates and yields would stay at or above zero, or close enough that code built for positive numbers would cope.
  • The path: prices would move continuously, so that stop-loss orders and margin calls would execute near their trigger levels.
  • The client: customers trading on margin could cover their own losses, so a broker’s exposure ended at the client’s account balance.

After the four assumptions we take the strongest objections to this framing. Then we set out what a chief data officer, chief risk officer or head of model risk can build so the same assumptions do not hide in current models. Those include the machine learning models now used in credit, fraud and treasury, which learn from calm periods as readily as value-at-risk did.

The scope is financial services broadly, and the four assumptions apply differently by type of firm. Retail foreign exchange brokers took the visible losses. Banks with currency desks, wealth managers with Swiss franc mortgages and structured products, asset managers with currency hedges and treasury functions at any firm with franc exposure faced the same assumptions in different forms. A credit union or community bank with no franc exposure still faced the number-range assumption, because negative rates in Europe pulled down yields worldwide and tested every system that priced deposits and loans. A broker-dealer faced the path and client assumptions in its margin lending. An insurer faced the history assumption in the asset models behind its investment portfolio. The analysis below is written for the firms that faced all four, but most institutions will recognize at least two.

The history: three years of a managed price

From September 2011 to January 2015, the euro-franc exchange rate could not fall below 1.20, because the SNB stood ready to buy unlimited foreign currency to defend it (SNB). The rate traded above the floor, often close to it. Its day-to-day volatility was low, and its downside was capped by policy.

What the calibration window saw

Most market risk models in 2015 calibrated to history. A historical value-at-risk model takes the last one or two years of daily price changes and asks how bad a day could be. A parametric model estimates volatility and correlation from a similar window. Margin models at brokers and clearing houses often did the same. On 14 January 2015, every one of those windows for the euro-franc rate sat inside the floor period.

The result was predictable in hindsight. The models reported low risk for long franc positions against the euro and for short positions in the euro against the franc, because the data had never shown a large move. The data was accurate. It described a world in which a central bank was absorbing the risk. The models had no way to know that the absorber could leave.

Why this is a data problem

Risk teams knew the floor existed. Traders talked about it daily. The problem was that the knowledge lived in people’s heads and in market commentary, not in the data that fed the models. The price series carried no flag saying it was recorded under a policy floor. A model calibrated on that series inherited the floor’s calm without inheriting the condition that made the calm possible.

A regime flag is a simple data element. For each series, it records whether the price was free, managed within a band, pegged or floored, and by whom. With such a flag, a model validation rule can refuse to calibrate on a managed period, or require a documented override. Without it, the calm period passes as ordinary history.

Stressed periods help only if the right one is chosen

Regulators had already asked banks to look beyond recent history. Stressed value-at-risk, introduced after the 2008 crisis, required banks to calibrate to a period of significant stress. But the stressed period was usually chosen for the portfolio as a whole, often the 2008 crisis itself. For the franc, 2008 was a period before the floor existed, and the franc had then strengthened sharply. A bank whose stressed period captured that move had some protection. A bank whose stressed period was chosen for other risks did not.

The implication for data teams is narrow and practical. Stress calibration needs to be chosen per risk factor as well as per portfolio, and the choice needs metadata. For each factor, the data should say which period best represents its free-floating behavior and why. That record is evidence an examiner can review.

Correlation carries the regime too

Volatility was one casualty. Correlation was another. While the floor held, the euro-franc rate barely moved, so the franc’s moves against the dollar tracked the euro’s almost exactly. A risk model estimating correlations from that period would treat franc and euro exposures as near substitutes. A portfolio long francs and short euros against the dollar would look almost hedged.

The floor’s end broke that relationship in a day. Positions that the model saw as offsetting became large, one-directional bets. Firms that hedged franc exposure with euros, or the reverse, found the hedge had been an artifact of policy. The data had shown a stable relationship because a central bank enforced it.

For data teams, this means regime flags should travel with derived data as well as raw prices. A correlation matrix estimated on a managed period inherits the regime. So does any factor model, any principal components analysis and any machine learning feature built from those series. Lineage tools that record which raw series feed which derived measures make this possible. Without lineage, the regime flag stops at the first transformation.

Documenting the calibration window

Most model documentation in 2015 stated the length of the calibration window, such as two years of daily data. Fewer documented what happened in that window. A useful practice is a short calibration note for each model run: the window, the main risk factors, any regime changes or managed periods inside it, and any overrides. The note takes minutes to produce if the regime register exists. It turns an invisible assumption into a visible one that validators and auditors can challenge.

Where the history assumption hides today

Managed prices did not end in 2015. Many emerging market currencies are managed within bands. Some commodity prices are set or capped by governments. Policy rates move in steps chosen by committees. Deposit rates at banks lag market rates by design. Each of these series produces calm data while the manager holds, and violent data when the manager stops.

Machine learning models face the same trap. A model trained on several years of deposit behavior during near-zero rates learned how depositors behave when rates do not matter. A fraud model trained during a period of stable payment patterns learned what normal looked like before a new payment rail changed it. The training data can be accurate and still describe a regime that is about to end.

The number range: code written for positive numbers

The SNB’s decision pushed its deposit rate to minus 0.75 percent and its Libor target range as low as minus 1.25 percent (SNB). The ECB’s deposit rate was already below zero. Denmark cut its certificate of deposit rate twice in the week of 19 January, to minus 0.35 percent by the end of the week (Danmarks Nationalbank). The ECB’s bond buying, announced on 22 January, put further downward pressure on euro area yields (ECB). Negative numbers were moving from an oddity to a regular part of European rates data.

Where negative numbers break things

Financial software makes assumptions about ranges, many of them never written down. Three kinds were exposed by negative rates.

The first is models. Many standard option pricing models for interest rates assume that rates follow a lognormal distribution, which cannot go below zero. Feed such a model a negative forward rate and it returns an error, a nonsensical value or nothing at all. Practitioners worked around this by shifting rates up by a fixed amount before pricing, or by switching to a normal model that allows negatives. Each workaround changed the meaning of the volatility numbers that traders quoted and that risk systems stored.

The second is code. Interest accrual routines, deposit pricing engines and loan systems were written on the assumption that interest flows from borrower to lender. A negative rate reverses that flow. Code that floors rates at zero, takes logarithms of rates, or treats a negative value as an input error can fail in ways that are not obvious, such as silently setting a rate to zero.

The third is data validation. Market data platforms check incoming prices against plausible ranges to catch bad ticks. A rule that rejects negative yields as errors will drop valid data, leaving models to run on stale values. Nobody notices until a reconciliation breaks or a valuation looks wrong.

Deposits, loans and customer-facing systems

Negative rates also reached retail and commercial banking systems, more slowly. Banks had to decide whether to pass negative rates to depositors, and if so to whom. Some chose to charge only large corporate and institutional depositors, often through fees rather than a negative interest rate, partly because their core banking systems could not post negative interest. A fee is a workaround that keeps the number positive in the system while the economics are negative.

Loans raised the opposite question. A floating-rate loan priced at a benchmark plus a margin could, in principle, produce a rate below the margin or below zero. Loan contracts written before 2014 rarely said what happened then. Some lenders applied zero floors to the benchmark. Others had to work out what their contracts allowed. Either way, the systems had to implement a rule that someone had decided, and in many cases nobody had decided it in advance.

The SNB’s own design added another data task for Swiss banks. Its negative rate applied only to sight deposit balances above an exemption threshold (SNB). Each bank had to calculate its threshold and its exposure, and manage balances around it. That is a small example of how a policy decision becomes a data and calculation requirement for every institution it touches.

What a range audit involves

The fix starts with an inventory. For each system that stores, validates or calculates with a rate, yield, spread or price, the team records whether it accepts negative values and zero, and what happens at the boundaries. The inventory covers models, code libraries, validation rules, database column types and downstream reports.

Testing follows. Each system gets a test set with negative values, zero and very small positive values, and the results are compared with an independent calculation. Failures go into a remediation plan with an owner and a date. The tests then stay in the release pipeline, because new code reintroduces the same assumptions.

This is unglamorous work, and it competes for budget with visible projects. It is also cheap compared with the alternative: discovering the boundary in production, during a market event, when the systems are under the most strain.

The same issue in commodity and credit data

Rates were the first large market to go negative, but they were not the last. Negative prices, spreads and basis values appear in energy, power and some credit markets. A data platform that handles negative rates correctly may still have commodity routines that assume positive prices. The range audit should cover every asset class the firm prices, including those beyond the one that triggered the review.

The path: prices that jump instead of move

The third assumption is about how prices get from one level to another. Most risk and trading systems assume continuity: a price moving from 1.20 to 1.10 passes through 1.19, 1.18 and every level between. Stop-loss orders, margin close-outs and hedging rules all depend on that.

What happened to the path on 15 January

When the floor went, the euro-franc rate did not glide. Bloomberg reported that the franc rose as much as 41 percent against the euro during the day (Bloomberg via swissinfo.ch). Liquidity disappeared. Alpari UK cited an extreme lack of liquidity when it announced its insolvency, and Deutsche Bank’s Autobahn platform briefly stopped quoting (Bloomberg via swissinfo.ch). A stop-loss order at 1.19 could not be filled at 1.19 because nobody would trade there. It was filled, if at all, far below.

That gap is the mechanism behind the negative client balances. A client with a margined position and a stop-loss assumed the loss was capped at the stop. The broker assumed the same. When the fill came far beyond the stop, the loss exceeded the client’s equity, and the excess became the broker’s problem (FXCM).

How the path assumption enters models

Value-at-risk at a one-day or ten-day horizon assumes positions can be adjusted within that horizon at prices close to the model’s distribution. Margin models assume they can close out a defaulting client within a set time. Hedging models assume rebalancing at observed prices. Each is a continuity assumption with a time scale attached.

The data that supports those assumptions is market depth and liquidity data, not prices alone. A risk team that stores only closing or mid prices cannot test how much of a position could be sold at what cost in a stressed market. Order book snapshots, quoted spreads and dealer quote counts are the data that would show liquidity thinning, and they are often the first to be dropped from storage because of their volume.

Gap scenarios as a data product

For a managed price, the right stress test is not a historical move scaled up. It is a gap: the price jumps from the managed level to an estimate of its free level, with no trading in between. Building that scenario requires an estimate of the free level, which is a judgment, and a decision about which positions and orders would execute where.

Such scenarios belong in the firm’s scenario library with documentation of their assumptions, an owner and a review date. They should be run against client positions, open orders and the firm’s own book. The output is a list of exposures that would breach limits or leave negative balances if the gap happened tomorrow.

Prices that stop arriving

A gap has a second effect on data: feeds stop. When a platform stops quoting, as Deutsche Bank’s Autobahn briefly did (Bloomberg via swissinfo.ch), downstream systems receive nothing. Many systems then use the last good price. During a gap, the last good price may be the old, managed level, hours out of date.

Stale data rules decide what happens next. A good rule flags a price as stale after a set interval, stops automated actions that depend on it and alerts a person. A weak rule carries the old price forward silently, so valuations, margin checks and risk reports show positions as safe while the market has moved far away. The difference matters most in the first hour of a gap, when decisions are made fastest.

Market data teams should define staleness thresholds by instrument and by time of day, record when each price was last updated, and show that age on risk screens. They should also decide in advance which sources take over when a primary feed stops, and test the switch.

Announced decisions and unscheduled ones

Central bank decisions usually arrive on a published calendar, and markets prepare for them. The SNB’s move came outside its usual quarterly assessment schedule, which added to the shock. Firms that hold larger buffers around scheduled events may hold smaller ones on ordinary days. A gap scenario does not care about the calendar. The firm’s exposure on an ordinary Wednesday has to be survivable if the regime ends that morning.

The client: borrowed exposure that passes losses upstream

The fourth assumption concerns who bears the loss. Retail foreign exchange brokers let clients trade currency pairs on thin margin, often with exposures many times their deposits. The business model assumed that client losses would stop at the client’s account balance, either because margin calls or stops would close positions in time, or because clients would pay any shortfall.

Where the assumption failed

Both mechanisms failed together. Stops filled far from their triggers, so losses ran past client equity. Many clients with negative balances could not or would not pay. FXCM’s statement put the resulting negative equity owed to the firm at about $225 million and warned of a possible capital breach (FXCM). Alpari UK, Global Brokers NZ and others took similar hits relative to their size (Bloomberg via swissinfo.ch).

For brokers, the client balance was a data assumption as much as a credit one. Client accounts were treated as self-contained. Risk aggregation looked at the broker’s net position after hedging with liquidity providers, not at the gross set of client accounts that could go negative together. A broker that hedged its net exposure could still be ruined if many clients on the same side of the trade went negative at once.

Why the clients were on the same side

Client positions were not random. Many retail traders had bet on the floor holding, buying euros against francs near 1.20 on the theory that the SNB would defend it. Such positions offered small, steady gains with what looked like a policy-backed limit on losses. That made them popular, and it made client positions highly correlated. When the floor went, a large share of clients lost together.

Concentration of client positions is a data signal. A broker that tracks the share of clients and of client margin on each side of each pair can see crowding build. When most clients are positioned the same way on a managed price, the gap scenario is the main risk, not a tail case.

The data needed to see it

The relevant data was the distribution of client margin and position size on each managed currency pair, combined with the gap scenario. That means aggregating, daily, each client’s position, margin and stop levels, and computing how many accounts would go negative under the gap and by how much. The output is a single number: the broker’s expected receivable from clients who cannot pay. Set beside regulatory capital, that number says whether the firm survives the gap.

The same logic applies beyond retail brokers. Banks that lent in francs to borrowers whose income was in other currencies, a practice found in parts of Central and Eastern Europe, faced a slower version of the same problem. Prime brokers financing borrowed positions at hedge funds on the wrong side of the franc faced it too. In each case the question is how much of the client’s loss can pass to the firm, under which scenario.

Margin set for the regime, not the volatility

The deeper fix is in how margin is set. Margin calibrated on observed volatility falls when a peg suppresses volatility, which invites larger positions exactly where the hidden risk is largest. Margin set for the regime ending stays high however calm the data looks. That choice costs business, because clients go where margin is cheaper. It is still the choice that keeps the broker solvent.

What clients were told

Disclosure was a further gap. Many clients understood a stop-loss as a guarantee. It is an instruction to trade at the next available price, which in a gap can be far away. Brokers’ terms usually said so, in small print. After January 2015, the gap between what clients thought and what the terms said became a dispute with each client who owed money. For data teams, the relevant point is that client-facing systems and statements should describe the risk of a gap clearly, using the same scenarios the risk function runs.

Where this framing is weakest

The strongest objection is that the SNB’s move was a genuine surprise, and no data discipline could have predicted its timing. That is correct. The argument here does not depend on prediction. A regime flag, a range audit, a gap scenario and a client-aggregation report do not tell a firm when a floor will end. They tell it what happens if it does, and whether the firm can survive it. Firms that ran such analysis could size their exposure without guessing the date. The same is true of the ECB’s decision a week later. Its timing was widely expected, but its scale and duration were not known until it was announced (ECB). A firm whose systems had already been tested on negative yields did not need to know the size of the program to know its systems would cope.

A second objection is that the losses were concentrated in retail brokers with thin capital, and banks came through largely intact. Banks did come through better, and their capital rules played a part. But the assumptions were present in bank systems too. Several were exposed more slowly, through negative rates in deposit and loan systems over the following years rather than in a single day. Concentration of the visible losses does not mean the hidden ones were absent.

A third objection is that regime flags and gap scenarios add judgment to models that are supposed to be objective. Someone has to decide which prices are managed and what the free level would be. Those decisions can be wrong. We accept that. A model that ignores the regime is also making a judgment, namely that the regime will hold forever. The choice is between explicit judgment, documented and reviewed, and implicit judgment, hidden in the calibration window.

A fourth objection is cost, and it is the one most likely to decide the matter inside a firm. Range audits across legacy systems, liquidity data storage and daily client aggregation are not cheap. For a large bank, the range audit alone can touch hundreds of applications. We think the cost is best judged against a single bad day. FXCM’s client receivable from one morning was about $225 million (FXCM). Most data programs cost far less than that.

The cost argument also cuts the other way over time. The range audit and boundary tests, once built, protect against every later move across zero, in any asset class. The regime register serves every managed price the firm will ever hold. These are fixed investments with long lives, which is the kind of spending a data function should find easiest to justify.

A fifth objection is that the problem is model risk, not data, and belongs to quantitative teams rather than data teams. Model risk management does own model choice and validation. But each of the four assumptions here depended on data that was missing or unlabeled: no regime flag, untested ranges, no liquidity history, no gross client aggregation. Model validators can only test what the data lets them see. Fixing the data is the precondition for the model work.

A sixth objection holds that the four assumptions overlap: a gap is partly a history problem, and client losses follow from gaps. They are connected, since the same event triggered all four. They are still distinct in the data they need and the people who own them. Calibration windows belong to market risk, value ranges to engineering, liquidity and staleness to market data, and client aggregation to credit risk. Treating them as one problem tends to leave each owner assuming another has it covered.

What data and risk leaders can build

The four assumptions point to four data products, each with an owner and a measurable output. None requires a new platform. All require someone to be accountable for them.

A regime register

A register lists every market data series the firm uses for risk or pricing, with its regime status, the authority managing it and the date of the last review. Model calibration routines read the register and refuse managed periods unless an override is documented. The register’s owner is usually the market data or chief data office, with input from economists and traders. The measure of success is coverage: the share of risk factors with a current regime status.

A boundary test suite

A shared test suite holds negative, zero and extreme values for every rate, price and spread type. Every pricing library, accrual engine, validation rule and report runs against it in the release pipeline. The suite’s owner is the engineering or quality function. The measure is the share of in-scope systems that pass, and the age of any open failures.

A gap scenario library

A library of gap scenarios covers every managed price the firm is exposed to, with the assumed free level, the reasoning and a review date. Scenarios run daily or weekly against positions, orders and client accounts. The owner is market risk, with model risk validating the methods. The output is a short report for senior management showing exposures under each gap.

A gross client exposure view

For firms that extend margin credit to clients, a daily view aggregates client positions, margins and stops by risk factor and computes potential negative balances under each gap scenario. The owner is credit or client risk. The output is the firm’s potential receivable from clients under stress, compared with capital. For a bank or broker-dealer, the same view should feed the credit limits set for each client segment, so that the people who approve margin terms see the gap exposure their approvals create.

Governance for machine learning models

The same four assumptions apply to machine learning models. Model inventories should record, for each model:

  • whether the training data was recorded under a regime that might end, and which one;
  • whether live inputs can fall outside the range the model saw in training, including negative values and zero;
  • whether the model assumes behavior changes smoothly when it could jump;
  • whether its output passes risk to a customer or counterparty who cannot bear it.

Validation should test each of those answers rather than accept them as stated.

Sequencing the work

The four products do not need to arrive together. The boundary test suite usually comes first, because it finds latent defects that could fail on any day and needs no new data. The regime register follows, since it depends on deciding which series are managed and takes economists’ time. Gap scenarios build on the register. The gross client view comes last for most firms, because it depends on clean client and position data that may sit in several systems.

Each step should end with a measured output reported to the risk committee: tests passed, series covered, scenarios run, exposures found. That reporting keeps the work visible when markets are calm and the budget pressure to cut it is highest.

What examiners and auditors look for

Supervisors reviewing model risk tend to ask whether a firm understands the limitations of its models and the data behind them. The four products give a direct answer. The regime register shows which data the firm treats as unrepresentative. The boundary suite shows which ranges have been tested. The scenario library shows how the firm sizes exposures that history does not show. The client view shows who bears the loss. Internal audit can test each against its documentation, which is easier than auditing a judgment that was never written down. In the United States, the banking agencies’ 2011 supervisory guidance on model risk management already expects firms to understand model limitations and the data behind them, and machine learning models fall within its scope.

With hindsight

Regulators moved first on the market risk rules. On 14 January 2016, almost exactly a year after the franc shock, the Basel Committee published a revised framework for market risk capital. It replaced value-at-risk with an expected shortfall measure of risk under stress, meant to capture tail risk. It replaced the single ten-day horizon with varying liquidity horizons for different instruments, to address the risk of sudden market illiquidity. It also tightened approval of internal models, allowing supervisors to withdraw approval desk by desk (Basel Committee). The framework was not written for the franc alone, but it addressed two of the four assumptions here: calibration to calm history and continuous paths. Its implementation dates slipped repeatedly in the years that followed and varied by jurisdiction.

European securities regulators addressed the client assumption for retail investors. From 1 August 2018, the European Securities and Markets Authority’s temporary measures on contracts for difference required providers to cap a retail client’s losses at the funds in the account. They came alongside limits on borrowed exposure that varied with the underlying and a margin close-out rule. Negative balance protection became standard for European retail clients in margined products. That moved the gap risk from the client to the broker by rule, which is where January 2015 had already shown it landed in fact.

The number-range assumption was tested again, more severely, in 2020. On 20 April 2020 the expiring U.S. crude oil futures contract settled below zero. CME Group had already prepared for negative prices and switched its options pricing model for affected energy products to the Bachelier model, which allows negative underlying prices, with effect from 22 April. The episode showed that code and models in commodities carried the same positive-number assumption rates had carried five years earlier.

The path assumption returned in March 2022, when the London Metal Exchange suspended nickel trading after an extreme price spike and voided trades executed that morning. The decision was contested in court. For risk teams, it showed again that a market can stop providing prices at exactly the moment models assume it will.

Negative rates themselves lasted longer than many expected in January 2015, then ended. The SNB kept its policy rate below zero until September 2022, when it raised it to 0.5 percent. The ECB’s deposit rate returned to zero in July 2022 and turned positive in September. In June 2025 the SNB cut its policy rate back to zero, and it held there into 2026. The boundary tests that firms built for negative rates remained useful, since the range of possible rates now plainly included both sides of zero.

The larger test of the following decade ran the other way. In 2022 and 2023, firms whose models and balance sheets had been calibrated to years of near-zero rates faced the fastest rate rises in decades. The bank failures of March 2023 in the United States turned in part on assumptions about deposit behavior and interest rate risk that had been learned in a low-rate regime. It was the same failure as January 2015 with the sign reversed: a calm history treated as a forecast.

What held up is the claim that these were data failures. Each later episode was a case of a model trusting data recorded under conditions that were about to change, or code assuming a range the world was about to leave. What did not hold up was the hope, common in early 2015, that the franc was a one-off. Regimes keep ending, and models keep learning from the years before they do.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • model risk
  • market data
  • negative interest rates
  • Swiss franc
  • stress testing
  • foreign exchange