FRTB – A Brief History (of time-series)

In the beginning was VaR

Value at Risk (VaR) can trace its roots back to 1990s and was developed as a way of aggregating risk positions across different product groups and asset classes. It is largely used by the Market Risk department (MRM) as a way of providing a single risk number to senior management on the amount the firm can lose during a bad trading day (worst in 100 trading days or fifth worse day in two years). VaR calculations tend to be run by the MRM department in overnight batches where they have to process risk data from several trading desks and run the simulations based on how the market has moved in the previous two years.

From a trader’s perspective, VaR is not relevant for managing their risk. The results are next day so not immediate and too high-level to be reconcilable back to what risk information they use to hedge, but managers liked it as it gives them a single number that facilitates conversations with the banks’ boards and the regulators.

In spite of the fact VaR has more caveats than most insurance policies (e.g. It is an estimate of loss based by looking in the rear mirror and might not reflect the risks in the future), the measure was taken up by the Risk Great and Good (Basel Committee) as the preferred methodology for calculating an appropriate capital charge. In 1996 the Basel Committee extended the current “Basel 1” standard approach framework to allow internal models based on VaR. The Basel Committee tasked the regulators to approve these new models. The regulators gave the banks the freedom to develop their methodologies and processes and provided a light-touch approach to reviewing the data, processes and systems of their VaR engines.

Throughout the 2000s the majority of banks built their own VaR engines, coinciding with a massive increase in the amount of risk warehoused. MRM shoe-horned as much of this new risk into their VaR engines as the alternative of calculating the capital under the original “Basel 1” standard rules would make these products uneconomic. During this period, Basel Committee brought out a new Basel 2 Accord to address other risk issues.

Did someone mention Caveats?

The once in a generation movements during financial crisis created daily losses that dwarfed the one-day-in-a-hundred VaR estimates, calling into question of whether the VaR had sufficient memory of stress periods. In 2009, the Basel Committee fast-tracked Stressed VaR or sVaR that covers the events of the financial crisis. The market risk capital charge now includes VaR + sVaR in what the Risk community now call the Basel 2.5 Accord.

From 2009-13 the regulators focused on improving the scope of risks that are included in the capital charges and captured in Basel 3 Accord, and also ensuring there is additional buffers for banks to cope with potential losses that arise from regulators running a program of stress testing.

Basel 2.5 and stress testing are significant initiatives for banks’ MRM departments and their regulators. The amount of capital now charged for warehousing market risk had increased substantially leading to banks taking on much less market risk compared to the previous decade. However, the financial crisis and the subsequent more invasive reviews of the banks’ operations continue to raise many gripes that needed resolution: –

Trading Book Banking Book. There was a suspicion that the motivation of the Front Office to move illiquid previously traded positions to the Banking Book was to reduce capital charge
Stressed VaR. The Basel committee introduced sVaR in a hurry. VaR + sVaR does not address the need for capital charges to capture events throughout an entire economic cycle. VaR has the habit of dropping during the quiet period just before the storm of a market correction
Model Variability. The regulators are unable to compare bank’s VaR calculations like-for-like as they use very different methodologies. When the regulators provided a portfolio for each bank to calculate VaR (known as Hypothetical backtesting), the difference in the results was material
Liquidity Horizon. VaR assumes that traders can access liquidity to exit their position in two weeks. The financial crises showed just how quickly liquidity could dry up, leading to reluctant and long-term holdings with no opportunity to exit at any price
Historical Timeseries. Banks use a large number of historical data on risk factors that do not change for long periods due to inactivity in the market. These gaps in activity go against the spirit of a daily VaR measure based on daily observation of activity in the relevant instruments
Risk Culture. There was a general frustration from the regulators that Front Office continues to be not that interested in using VaR as an integral part of their risk management toolset even though MRM administer formal limits as part of the risk governance framework

Addressing the Gripes

These persistent gripes prompted the Basel committee to introduce a new set of requirements for market risk capital called the Fundamental Review of the Trading Book (FRTB). The risk community uses FRTB to describe the new body of regulation whereas logic would suggest Basel 4. The risk community got there first, so its FRTB for the forceable futures.

FRTB therefore reads as a list of instructions that address each of the issues raised above:-

Trading Book Banking Book. Movement between Trading and Banking Book is now severely restricted and has to go through a formal governance process where MRM has to make sure that the capital does not fall
Stressed VaR. The VaR lookback period has increased from 1-2 years to ten years. The sVaR calculation remains and the calculation restricts the netting benefit of hedge positions that post gains during stressed periods
Model Variability. Basel has provided a pre-canned methodology for aggregating market risk that all banks will adopt so that regulators can conduct reviews across peer groups. This standard methodology acts as a fall-back to the banks’ current internal models
Liquidity Horizon. Extension of liquidity horizons that reflect the access of each product to liquidity during periods of stress. Liquidity horizons now range from two weeks to a year.
Historical Timeseries. Introduce a requirement that for risk factors used in the VaR engine, their historical data needs to pass a number of quality checks including a minimum number of observations and length of inactivity
Risk Culture. Model approval moves away from being a negotiation between the regulators and MRM at a firm level. Regulators will grant approval to individual desks that present themselves as standalone business units where heads must show that they are in control of all of their risks including active participation in how the VaR engine processes their risk

FRTB Alphabet Spaghetti

The introduction of FRTB has triggered the introduction of an alphabet soup of new acronyms:-

RWA. Risk Weighted Assets. This allows market risk positions to aggregate alongside assets from other activities such as loans. The need to weight assets gives rise to VaR and all capital methodologies
VaR, sVaR, and ES. FRTB will replace VaR and sVaR with Expected Shortfall or ES that takes the average of extreme losses in the VaR engine rather than a single point. The reason for the shift is of great interest to a small number of academics.
IRC and DRC. Incremental Risk Charge. This simulates losses on credit-traded instruments brought on by downgrades and defaults of their issuers. This is being replaced by Default Risk Charge or DRC that captures losses from defaults only but now includes equity positions
SA, SBA. Standard Approach. This is the new pre-canned calculation that all banks have to do the same calculation on all of their positions. Part of SA is the Sensitivity-Based Approach or SBA that describes in detail how banks aggregate their risk information and is a conservative alternative to VaR. Some regulators use the term Market Risk Standard Approach or MRSA to refer to positions that banks have chosen not to include in VaR and calculate capital based on the basic Basel 1 charge
IM, IMA. Internal Model. Models such as VaR/sVaR and ES that banks build and regulators approve. Internal Model Approach or IMA is the alternative to Standard Approach and will attract less capital but requires the regulators to approve their models
RNIV, NMRF. Risk not in VaR. Risk that arise from usually client activity in one-off bespoke trades and unlikely to be included in a VaR calculation. Non-Modellable Risk Factors or NMRF extends RNIV to include risk factors that are in the VaR calculation but do not pass a number of quality tests

Unlike previous approval processes that have been a more mathematical discussion with MRM around the distribution of results that the methodology produces – commonly called “Fat Tails” and is a way of asking if enough large losses are below the VaR estimate.

For FRTB the will ask each desk head questions around the desks capability to identify, measure, monitor all of their risk. The regulators will look how each desk can articulate what is the business they are, what products do they offer, how they intend to make money and how much financial resources do they intend to consume. The regulators will expect that each desk have their own governance framework as well as an understanding how they plug into corporate governance such as New Product Approval, Model Governance and VaR/ES methodology governance and Audit Review.

From a measurement perspective, the regulators will expect to review a year’s supply of daily monitoring of the following:-

Risk Factors. List of risk factors used for pricing and which the Front Office pass to MRM. Which risk factors has MRM classified as modellable, or are NMRF. Which risk factors are capitalised outside of VaR (RNIV)
Risk Reports. Daily (and possibly intra-day) risk reports using the Front office risk factors and an equivalent but simplified report using MRM risk factors
Eligibility Tests. Detailed comparisons between the Actual P&L and the Front Office Attributed P&L and MRM VaR-based P&L, check that they pass the tests
Trades. Daily monitoring of transactions viewed by risk factors (used in NMRF) as well as inventory turnover for aged reporting
VaR Backtesting. Detailed breakdown of the losses that breach the VaR, whether this is a genuine tail event or that it is sufficiently covered by capital add-ons due to remarking of risk factors outside of the modellable set
Concentrations. Risk sensitivity reports against MRM limits, concentration to single (or multiple) name default, pricing model usage

The Sting – Eligibility Tests

Experience of regulatory approval for internal models for market, credit and counterparty risk suggest the process takes 12-18 months from initial submission through to final approval and switch over. With desktop approval increasing the workload of the regulators, audit will do much of the legwork to alleviate the pain and banks will have to be quite selective on which desks they go for IMA.

Once the regulators have granted Internal Model approval, each desk must continuously monitor that VaR retains its status of being a one-day-in-a-hundred against the actual loss. If the actual loss is higher than the VaR on three or more days in a 12-month window, the regulators can ask questions on whether the VaR is capturing enough risk. If the P&L consistently breaches the VaR then the regulators will conclude that the portfolio is not sufficiently capitalised and the desk reverts to using the more conservative Standard Approache.

This “VaR Backtesting” was a requirement for Basel 2.5, however the analysis would only kick in a when a loss breached the VaR, so not really in the spirit of continuous monitoring. The explanations also became so reliant on “it’s another once-in-a-lifetime event”, so started to echo the Queen of Hearts. The banks needed to answer whether the VaR is a good predictor of a wider set of losses, in particular as markets go through a correction over weeks rather than one day.

At its core, the VaR is a list of P&L estimates from simulating the performance of the current portfolio against historical market movements. The Basel Committee introduced a new eligibility test that compares the actual P&L that the finance directors have signed off with the approximate P&L used in the VaR engine.

The VaR process involves taking a summary of the risk used by each trading group and therefore lacks the detailed view a trader would recognise. This “normalisation” is a natural consequence of collecting risk data from multiple systems and find ten years of historical data for the risk factors. Not everything has been around for 10 years; so much risk is proxied to more established risk factors. This leads to deviation between the actual P&L and an estimate of the P&L used in the VaR engine. If this deviation hits a threshold, the regulators will move the desk back to the Standard Approach, leading to an increase in capital charged.

Analysis in 2016 conducted by ISDA suggested that upwards of 90% of trading desks would fail the P&L eligibility test and therefore the banks have embarked on an exercise to improve the performance of their VaR engines, in particular increasing the number of risk factors so that it can capture more complex market movements.

Modellability? Smodellability!

The eligibility tests has prompted the banks to improve the granularity and hence accuracy of their VaR engines. The introduction of a capital add-on based on the concept of Non-Modellable Risk Factors unfortunately punishes those banks who have increased the number of risk factors in their VaR engines.

The regulators highlighted that when an instrument stops trading for a period, the historical timeseries flat-lines until someone posts a trade and the level for that risk factor jumps to reflect the new price. During this quiet period, the P&L is zero and the VaR for that risk factor trends down as more zero returns join the historical timeseries. This points to low liquidity and so the regulators require the banks to adjust the liquidity horizons to reflect evidence of longer gaps between observations.

The regulators also saw excessive gaps as evidence of under-capitalisation and hence decreed that the banks should not model them in their VaR engine. The FRTB text describes them as “Non-Modellable Risk Factors” or NMRF. Banks can either take them out of the VaR engine and classify them as Risk Not in VaR or keep them in and apply a capital add-on.

To retain modellability, banks must show for each risk factor regulators evidence of at least 24 trades or committed quotes per year with no more than a month between any 2 observations. For each NMRF, MRM has to develop a sufficiently severe stress scenario that they use to apply a capital add-on.

At first glance the test looks quite modest, however the number of risk factors used by Tier 1 is several 100,000s as each curve is bucketed into a large number of risk factors. The number of NMRF has increased thanks to the work around the eligibility test. Banks indicate that NMRF add-on will be 30% of the overall FRTB capital charge and therefore represents the bulk of the expected 60% uptick in market risk capital.

FRTB – will it every happen?

Since 2016 FRTB has been the bridesmaid to IFRS9 and MIFID II which meant that development work as been scratchy.

For 2019/20, the banks will need to deliver the standard approach (SA or SBA) for all books as well as start the work on the desks highlighted for internal model approval (IMA). Key to delivery of the SA are detailed information on the drivers of the capital charge so that Front Office and their Infrastructure areas can continue to refine the data quality and their hedging strategies. For IMA, the desks need to be complete by end of 2020 to leave 2021 to build out a year of quality results. This is a tall order so it is critical that as small a set of desks is highlighted as possible. The other desks can follow later.

Back to 2007

Banks will need to extend the historical timeseries back to 2007,:-

Monte Carlo now Bust. Banks that used a Monte Carlo Simulation benefited from the fact the maths conveniently smoothed over a number of challenges such as granularity experienced in Historical Simulation VaR (or Hist Sim VaR or just VaR). With FRTB, these banks are required to move their methodology across to the industry norm, which means they need to source a much greater set of historical timeseries. The historical timeseries is also required to calibrate the correlations required to simulate clusters of defaulting names
Move to Expected Shortfall. ESis sensitive to the extreme outliers so having analytics that enable banks to improve the quality of the timeseries in particular cleaning out erroneous jumps will reduce level and variability of the results

Non-Modellable Risk Factors

For each of the their risk factors used in their VaR engine, each bank will need to keep store of trading activity to make sure the conditions are met to continue being modellable. This monitoring is continuous and it is unlikely that any bank will be active in all risk factors all the time, so they will need to access details of trades executed away from them in order to fill the gaps in their own trade activity. Banks will need to have:-

Volume. Sufficient volume of transactional and committed quotes that the bank can aggregate alongside their own deals in order to pass the two NMRF tests With 10,000s of risk factors at risk of being either side of modellability, it is critical that banks know where they are at risk of additional capital charge and what instruments they need to trade
Early Warning. Trigger warning when the gap between transactions approaches a month. Once the gap has breached one month, the risk factor will drop into NMRF until the gap drops off after a year
Stress. Development and signoff of a stress scenario for each NMRF that needs to be at least as severe as what would be implied by VaR-type movement. Evidence that different approaches have been considered and analysis conducted
Liquidity Horizons. Mapping of risk factors to a minimum liquidity horizon as outlined in FRTB. A failure to have the right liquidity horizons will lead to a higher capital charge through over-statement or though non-approval thanks to under-statement. Adjust liquidity horizons if the maximum gap in the historical timeseries is longer than the minimum

Governance Landscape

The governance landscape for FRTB is considerably more complex and detailed than for Basel 2.5. Highlights include

Desk Level Approval. Each desk will need to maintain live documents and management information packs around; business strategy, budget and compensation mechanism, trader mandates, risk and flash PL, VaR breakdown and PL attribution, vendor fees and costs, Funding costs, CCP and Collateral reports, capital drivers, new client business, operational and processing reports. All have to go through a governance workflow to evidence of detailed review, challenge and actions with responsibilities
Micro Decisions. Banks build VaR on a mountain of micro decisions with each one needing to go through a governance process to evidence sufficient oversight including justifications, alternative approaches, independent review and impact analysis. These include changes in; timeseries sourcing methodology, timeseries gap filling, mapping risk to risk factors or proxy, stress scenarios for NMRF
Monitoring. Banks are required to monitor to ensure continued compliance with eligibility test and trade activity for NMRF. The banks need also to monitor; quality of each historical timeseries against metrics, inventory turnover, VaR back testing. All output needs to enter a workflow so that the bank can evidence that the reports have been received and that actions are being tracked. Banks need to avoid losing track of all of the items that need monitoring, who can see the output, who can assign action items and who can sign-off remediation efforts.

Each desk will need a particular slice of the data whether it comes from central monitoring or one that is local to the product set or region. Banks will nead an industrial strength solution just to keep track of all of the data items. The following monitoring is required for FRTB and the list is the edited highlights: –

End of day risk, P&L and market data signoff
Risk factor taxonomy and classification (Modellable, NMRF, RNIV)
Risk factor proxy rules that map Front Office risk factors to MRM risk factors
Risk sensitivity normalisation and tenor bucketing
Time-series data quality, review of abnormal jumps
Transactional data and inventory turnover
Gaps in risk factor time-series vs minimum liquidity (NMRF)
Number of NMRF, stress scenarios and last time checked for severity
Trading Book to Banking Book transfer workflow
Instrument to risk factor mapping (NMRF)
P&L eligibility tests
VaR back-testing and breach commentary