Synthetic market data generation: approach

Asset class	ORE quote types	Primary technique	Notes
OIS discount curves (EUR-ESTR, USD-SOFR, CHF-SARON, GBP-SONIA, JPY-TONAR)	IR_SWAP/RATE (1D index), OI_FUTURE/PRICE	GMM on par swap rate vectors	One GMM per currency-index pair; PCA (3–5 components) recommended for curves with >10 tenor points
Projection / IBOR curves	IR_SWAP/RATE (3M/6M index), BASIS_SWAP/BASIS_SPREAD	GMM (same fitting as OIS)	Basis spreads included in the rate vector or fitted as a separate low-dimensional GMM
Money-market / short end	MM/RATE, FRA/RATE, MM_FUTURE/PRICE, IMM_FRA/RATE	GMM or deterministic term-structure interpolation	Derive from the par swap GMM using standard bootstrap; do not generate independently
Zero / discount	ZERO/RATE, DISCOUNT/RATE, ZERO/YIELD_SPREAD	Derived	Bootstrap from IR_SWAP/RATE using QuantLib PiecewiseYieldCurve
Cross-currency	CC_BASIS_SWAP/BASIS_SPREAD, CC_FIX_FLOAT_SWAP/RATE	GMM or zero-basis static	Either include in the joint IR GMM, or set to zero for a pure-CIP environment

GMM fits on par rate levels (or first-differences for non-stationary regimes). PCA-preprocessing is mandatory for full-curve GMM with more than 10 tenor points: fit the GMM on PCA scores, sample in PCA space, reconstruct to par rate space, then bootstrap to a zero curve with QuantLib PiecewiseYieldCurve. Negative-rate environments (CHF, EUR pre-2022) require careful stationarity checks on the GMM components.

Asset class	ORE quote types	Primary technique	Notes
FX spot	FX/RATE	GMM on log-returns; reconstruct level	Ideally in the same joint GMM as IR; enforces IR-FX correlation
FX forwards	FXFWD/RATE	Derived only via CIP	Never sample from a GMM; see constraint 1
Cross-currency basis	CC_BASIS_SWAP/BASIS_SPREAD	GMM (small-dimensional) or zero	Add as a spread on top of CIP-derived forward

FX forwards are a derived quantity. After generating IR curves (step 1) and FX spot (step 3), every FXFWD/RATE tenor is computed as FX_spot × DF_for(T) / DF_dom(T) using QuantLib YieldTermStructure::discount(). Sampling FXFWD/RATE independently from a second GMM draw is prohibited: it breaks CIP, invalidates every cross-currency swap NPV, and makes delta hedging results meaningless. See 49687AD7 for the full CIP derivation.

Asset class	ORE quote types	Primary technique	Notes
Equity spot	EQUITY/PRICE	GMM on log-returns; reconstruct level	Include in joint GMM with IR and FX if cross-asset correlation matters
Dividend yield curve	EQUITY_DIVIDEND/RATE	Gap — static + noise	Flat dividend yield from ORE examples ±0.3% uniform noise per scenario
Equity forwards	EQUITY_FWD/PRICE	Derived only via carry	EQUITY_SPOT × exp((r − q) × T)

Equity forwards are derived after equity spot, IR curves, and dividend yield are all available. The dividend yield is a gap item (no paper generation method); a static flat yield with small random perturbation is the interim mitigation and is sufficient for most scenario generation purposes. QuantLib BlackScholesMertonProcess takes the dividend yield as a handle alongside the risk-free rate; the forward is consistent by construction when these inputs are used without modification.

Asset class	ORE quote types	Primary technique	Notes
CDS par spreads	CDS/CREDIT_SPREAD	Gap — historical sampling with date offset, or Nelson-Siegel (level + slope) with noise	No paper provides a validated generation procedure
Hazard rates	HAZARD_RATE/RATE	Derived from CDS spread + recovery	λ ≈ spread / (1 − R)
Recovery rates	RECOVERY_RATE/RATE	Gap — static (40% senior, 20% sub) ±5% uniform	Bounded [0,1]; GMM is unsuitable
Base correlations	CDS_INDEX/BASE_CORRELATION, INDEX_CDS_TRANCHE/BASE_CORRELATION	Gap — static ORE example values ±2% with monotonicity preservation	Derive from index CDS spread long-term
Rating transitions	RATING/TRANSITION_PROBABILITY	Gap — static matrices from ORE examples	Markov chain model required; out of scope

Credit is a full-coverage gap for this sprint. The interim mitigation for all credit types is to use historical ORE example data with small random perturbations that preserve the CDS-recovery constraint (fix recovery first, generate spread second). The proper solution — sector-level GMM on log-CDS-spreads per rating bucket with a copula for name-level correlation — is a Phase 4 item.

Asset class	ORE quote types	Primary technique	Notes
Commodity spot	COMMODITY/PRICE	GMM on forward curve vector (front contract = spot for energy)
Commodity forward strip	COMMODITY_FWD/PRICE	GMM on full tenor vector	Energy: deseasonalise before fitting, reapply seasonal shape after
Commodity options	See option vol surfaces	SSVI parameter dynamics	Treat like equity options; the commodity forward strip sets the forward each expiry is struck against
CPR / prepayment	CPR/RATE	Gap — static	Only needed for BalanceGuaranteedSwap; skip for initial scope

For precious metals, the forward curve must satisfy the carry relationship COMMODITY_FWD ≈ COMMODITY_SPOT × exp(r × T) (convenience yield ≈ 0 for gold, silver). Generate the spot and full IR curve first, then derive or check the forward strip post-generation. For energy, the forward curve is generated directly as a vector by the GMM; the front contract is the spot by definition and no separate derivation is needed.

Asset class	ORE quote types	Primary technique	Notes
ZC inflation swap rates	ZC_INFLATIONSWAP/RATE	GMM on ZC rate vectors (one GMM per index: EUHICP, EUHICPXT, AUCPI)
YY inflation swap rates	YY_INFLATIONSWAP/RATE	Derived from ZC curves via standard ZC-to-YY identity	Do not generate independently
Seasonality	SEASONALITY/RATE	Deterministic: ORE example factors or historical CPI seasonal adjustment	12-month product must ≈ 1.0
Inflation vol surfaces	ZC_INFLATIONCAPFLOOR/PRICE, ZC_INFLATIONCAPFLOOR/RATE_NVOL, YY_INFLATIONCAPFLOOR/PRICE, YY_INFLATIONCAPFLOOR/RATE_NVOL	Gap — static ORE example values × proportional lognormal scaling factor	None of the three papers covers inflation vol surfaces

YY swap rates must be derived from ZC curves — not generated independently — using the standard identity that converts the ZC compounding ratio at each annual period to a year-on-year rate. The real rate implied by the ZC inflation curve and the nominal IR curve must remain positive for the scenario to be economically coherent; post-generation checks flag violations.

Vol surface	ORE quote types	Pillar generation	Expiry interpolation	Strike/smile generation	Notes
Swaption cube (ATM)	SWAPTION/RATE_NVOL, SWAPTION/RATE_LNVOL	GMM on ATM normal vols with PCA (10–15 components for full cube)	Frozen pool (contingent on SABR-to-mixture conversion; see OD-3 in open design decisions)	GMM on SABR smile parameters (α, ρ, ν) per cell	Normal vols throughout for EUR, CHF; apply consistent model choice across swaption and cap/floor for same currency
Bond option vol	BOND_OPTION/RATE_LNVOL	Derived from swaption vol via duration mapping	Inherited from swaption cube	Duration-based approximation	Derive, do not generate independently

ATM swaption vols are generated by GMM on the PCA-compressed ATM vol cube. The frozen pool provides expiry-direction interpolation once the SABR-to-mixture conversion gap is resolved; until then, skip the frozen pool and write directly to SWAPTION/RATE_NVOL quote keys. Bond option vols are derived from the swaption cube using the ORE-standard approximation BOND_OPTION/RATE_LNVOL ≈ SWAPTION/RATE_LNVOL at (bond option expiry, modified duration × rate sensitivity).

Vol surface	ORE quote types	Pillar generation	Expiry interpolation	Strike/smile generation	Notes
Cap/floor cube	CAPFLOOR/RATE_NVOL, CAPFLOOR/RATE_LNVOL, CAPFLOOR/SHIFT	GMM (same PCA approach as swaption)	Frozen pool (contingent on OD-3)	GMM on smile shape	Shift must be computed from IR curve before vol generation (constraint 5); normal vol for negative-rate currencies

The CAPFLOOR/SHIFT for each currency must be derived from the IR curve generated in step 1: bootstrap the zero curve, identify the minimum forward rate at each tenor, and set shift ≥ |min_forward_rate| before populating cap/floor vol quote keys. This must happen at every time step when evolving the environment forward, not just at initial generation.

Vol surface	ORE quote types	Pillar generation	Expiry interpolation	Strike/smile generation	Notes
Equity option	EQUITY_OPTION/RATE_LNVOL, EQUITY_OPTION/PRICE	SSVI ATM total variance per expiry, struck against the equity forward (step 7)	Automatic: SSVI is a surface	SSVI skew and curvature; quotes by strike or moneyness	Free of static arbitrage by construction
FX option	FX_OPTION/RATE_LNVOL	SSVI ATM total variance per expiry, struck against the FX forward (step 4)	Automatic: SSVI is a surface	SSVI skew and curvature; ATM, RR and BF quoted in the pair's convention	Strikes solved from delta under the pair's delta and ATM conventions
Commodity option	COMMODITY_OPTION/RATE_LNVOL, COMMODITY_OPTION/RATE_NVOL	SSVI ATM total variance per expiry, struck against the commodity forward strip	Automatic: SSVI is a surface	SSVI skew and curvature	Seasonal/structural complexity is upstream in the commodity forward curve

The three surface types share one method: an SSVI surface whose parameters evolve as state processes. It is set out here for FX, the first asset class to use it; equity and commodity surfaces differ only in how they quote.

Asset class	ORE quote types	Primary technique	Notes
Pairwise correlation	CORRELATION/RATE (IR CMS-CMS, FX, equity)	Derived from jointly generated return vectors	Well-defined only if asset classes share a joint GMM

Correlation is not generated directly. It is derived as the Pearson correlation coefficient of the jointly sampled return pairs from the GMM draw across asset classes. This requires IR, FX, and equity to be included in the same joint GMM (or linked via a copula layer). If asset classes are generated with independent GMMs, all pairwise correlations are zero by construction and must be injected separately. The joint GMM architecture — block-diagonal, hierarchical, or copula-based — is open design decision OD-1.

GMM generates P-measure data: it is fitted to historical distributions and produces snapshots that describe what markets have done in the past. QuantLib's option pricers require Q-measure inputs: implied volatilities that are consistent with no-arbitrage option prices. A P-measure vol surface is not guaranteed to satisfy the calendar-spread or butterfly conditions (see 3F7A2C91), and feeding a violated surface into LocalVolSurface or ORE's local-vol pricing engine produces NaN or financially meaningless results. The P/Q distinction is not a theoretical concern — it causes concrete calibration failures in ORE. See A3F7C2E1 for the full measure distinction.

The surfaces of these three asset classes are generated directly in the Q-measure. The spot and forward paths are P-measure, from the GMM with its GJR-GARCH variance; the surface is a separate set of state processes, so it is never computed from a P-measure distribution and needs no conversion from one. Its Q-measure property is the one QuantLib and ORE need: every surface the generator emits is free of calendar-spread and butterfly arbitrage, because the SSVI parameters never leave the region where Corollary 4.1 of Gatheral and Jacquier holds, and every strike is solved against the forward of the same generated scenario. The method is in option vol surfaces.

The SSVI method covers the equity, FX and commodity option surfaces. The path for swaption and cap/floor surfaces is:

1. GMM step (P-measure): GMM on historical ATM normal vols (with PCA pre-processing) generates pillar ATM vols. These are P-measure statistics (what ATM vols have looked like historically).
2. SABR calibration step (Q-measure calibration): Treat the GMM-generated ATM vol as the market-observed ATM vol for each (expiry, tenor) pillar. Fit a SABR model to the ATM vol plus the smile shape (sampled from a separate GMM on SABR parameters α, ρ, ν). The SABR calibration enforces the no-arbitrage constraints implicit in the SABR model for each pillar smile.
3. Frozen pool step (Q-measure interpolation): Convert each SABR pillar smile to a Gaussian mixture density (fitting a normal mixture to ∂²C_SABR/∂K²) then apply the frozen pool for expiry-direction interpolation. This step is contingent on resolving OD-3 (the SABR-to-mixture conversion recipe). Until OD-3 is resolved, skip the frozen pool and write the ATM vol directly to SWAPTION/RATE_NVOL keys without smile or cross-expiry arbitrage enforcement.

For non-derivative instruments — FX/RATE, IR_SWAP/RATE, EQUITY/PRICE, COMMODITY_FWD/PRICE, ZC_INFLATIONSWAP/RATE — there is no P/Q distinction to resolve. A spot FX rate is a market observable, not a derivative price. The GMM-generated par swap rate or FX spot level is a direct input to ORE's curve bootstrapping infrastructure and requires no Q-measure adjustment. The P/Q distinction matters only for option surfaces (implied vols), because option prices are model-dependent and must be consistent with no-arbitrage.

The GMM is fitted to daily changes in each variable — first-differences of par swap rates, log-returns of equity and FX spots — not to levels. Sampling from the GMM gives a daily change vector Δx. The next day's state is constructed by applying Δx to the previous state:

• Par swap rates: r_{t+1} = r_t + Δr (first-difference GMM).
• FX and equity spots: S_{t+1} = S_t × exp(Δlog(S)) (log-return GMM).
• Inflation ZC rates: π_{t+1} = π_t + Δπ (first-difference GMM).

The initial state at t=0 can be a single ORE example snapshot (e.g. the Products example marketdata.csv) or a sampled GMM level.

After each daily change draw, before advancing to t+1, re-derive all dependent quantities in dependency order:

1. Re-derive FXFWD/RATE from new spot + new IR curves (CIP).
2. Re-derive EQUITY_FWD/PRICE from new spot + new IR + static dividend yield.
3. Re-compute CAPFLOOR/SHIFT from new IR zero curve minimum forward rate.
4. Advance the SSVI state of each option vol surface, and solve its quotes' strikes against the new forward.

The constraint table must be re-verified at each step.

At the end of each market day, the day's spot return ε_t updates the GARCH conditional variance:

σ²_{t+1} = ω + α·ε²_t + γ·ε²_t·𝟙{ε_t < 0} + β·σ²_t

σ²_{t+1} sets the size of the next day's GMM shocks. This is the state-carrying mechanism for the volatility of spot: the GARCH variance σ²_t persists between steps rather than being re-initialised from a GMM draw. The vol surface carries its own state, as option vol surfaces describes, and reads none of the GARCH variance.

The GMM fitted on daily changes produces i.i.d. draws. I.i.d. draws destroy temporal autocorrelation: consecutive sampled vols are uncorrelated. To preserve volatility clustering (high-vol periods followed by high-vol periods), the GARCH variance state σ²_t must be carried between steps. The GARCH mechanism provides the memory: a large return shock at t inflates σ²_{t+1} via the α and γ terms, which decays over subsequent steps with speed 1−α−β. Do not re-initialise σ²_t from a GMM draw at each step — that would discard the clustering and produce flat-vol, unrealistic scenarios.

Generation frequency — how often each asset class produces a new data point — is an independent, configurable parameter per asset class group. Different market variables update on different time scales and with different intraday patterns:

• FX spot ticks continuously during active market hours; quiet periods (overnight, weekends) see little movement. In a simulated scenario, FX spot may tick every 5 minutes during a simulated "active" window and every hour during a "quiet" window.
• Equity spots follow exchange hours; pre-market and post-market regimes have distinct volatility profiles.
• Interest rate curves update daily at standard market close times. Intraday updates are unusual unless simulating a stressed scenario.
• Volatility surfaces update less frequently than the underlying: a vol surface used for daily pricing might be regenerated once per business day even when the underlying spot ticks intraday.
• Credit spread curves and commodity forward curves typically update once daily or on material events.

This has three design implications:

Tick clock per asset class. Each asset class group carries a tick clock defined by an inter-tick interval distribution. Supported distributions:

Mode	Description	Parameters
Fixed	Deterministic interval	interval (seconds, minutes, days)
Poisson	Exponential inter-tick intervals — memoryless arrival	λ (mean ticks/unit time)
Regime-switching	Two-state (active/quiet) intensity — different λ per state	λ_active, λ_quiet, state-transition matrix
Hawkes	Self-exciting: a burst of ticks increases the probability of further ticks	μ (base rate), α (excitation), β (decay)

The Hawkes process is the natural model for "high-frequency generation, some quiet periods" — a cluster of FX spot ticks excites further ticks, reproducing the intraday burst patterns seen in live FX markets. For scenarios not requiring intraday resolution, Fixed (daily) is sufficient and computationally cheapest.

The mathematics of the Poisson and Hawkes modes is in Poisson Process and Hawkes Process. How a mode reaches the code, and the parameters it takes, is in Tick arrival.

Derived-quantity update rule. When a source asset class ticks, all derived quantities that depend on it must be updated immediately:

• FX spot tick → re-derive FX forwards (CIP) for all tenors.
• Equity spot tick → re-derive equity forwards (carry); add the move to the day's return that updates the GJR-GARCH variance at the day's end.
• IR curve tick → re-derive FX forwards, equity forwards, CAPFLOOR/SHIFT.

Derived quantities do NOT have their own independent tick clocks: they tick exactly when their inputs tick.

Cross-asset tick skew. When asset class A and B have different tick rates, there will be moments where one has ticked and the other has not. The generation engine must store the last-seen value for each asset class independently and construct a consistent snapshot on demand by combining the latest value of each class. This "mixed-vintage" snapshot is valid as long as all cross-asset constraints are re-verified at snapshot construction time.

Parameter	Practical range	Notes
Step length	Configurable per asset class (see tick frequency subsection above)	IR curves typically daily; FX spot can be sub-minute
Scenario horizon	1–5 years for most use cases	Intraday resolution increases step count proportionally
Regime conditioning	Base / stressed	For stressed scenarios, pin the GMM's component draw to the high-volatility component using the GMM's closed-form conditional: =P(x	component k) ∝ w_k · N(x; μ_k, Σ_k)=

Stressed scenario conditioning is a direct use of GMM's closed-form conditional distribution. No re-fitting is required: select the component index k with the highest unconditional variance and fix the latent component draw to k throughout the scenario.

Every technique in this document (GMM, GJR-GARCH, SSVI parameter dynamics, frozen-pool interpolation) is descriptive: each is fitted to historical distributions and reproduces their statistical shape, including the multi-component GMM's distinct "regimes" (steep/inverted curve, risk-on/risk-off). Fratini (2026) draws a sharp distinction between a descriptive model — one that organises past observations — and a generative model, which captures why a relationship holds and the conditions under which it breaks; a model can be "perfectly calibrated to the data and completely blind to the process generating it." His worked example is a risk-parity strategy that backtested well across equities, bonds, and commodities but never modelled the simultaneous-deleveraging feedback loop that inverted those correlations in March 2020 — the correlation matrix used to size the strategy was descriptively accurate for every regime seen in training, and still wrong for the regime that mattered.

The same caveat applies directly to stressed scenario conditioning above: pinning the GMM draw to its highest-variance component only reproduces stress episodes that already exist, in that shape, in the training window. It does not model the mechanism that produces a stress event (forced deleveraging, correlation flips, liquidity feedback), so a stress scenario outside the training distribution's component structure — or a correlation breakdown between asset classes that were never jointly stressed in history — will not be represented, however well each individual asset class's marginal distribution is calibrated. This is a known gap, not a defect to fix in this sprint: the mitigation (a true generative, mechanism-level stress model) is out of scope here and would need its own design work if higher-fidelity tail scenarios become a requirement.

Source: Alessio Fratini, "Have You Modeled the Market, or Only What It Has Shown?", LinkedIn Pulse, 2026 — https://www.linkedin.com/pulse/have-you-modeled-market-only-what-has-shown-alessio-fratini--su5ye

The control panel provides a dedicated screen in ORE Studio giving users full control over the generation process. It is not a configuration file editor — it is an interactive surface with immediate visual feedback.

Asset class and name selection:

The control panel provides per-group toggles for each asset class group (IR, FX, equity, credit, commodity, inflation, vol surfaces, correlation). For each active group, the panel provides a name selector: currency pairs for FX, currency and index combinations for IR (EUR-ESTR, USD-SOFR, CHF-SARON, etc.), ticker or index names for equity and commodity, inflation index for inflation (EUHICP, EUHICPXT, AUCPI). Selection drives which quote-key families are populated in the output.

Generation technique overrides:

The control panel provides per-group technique selectors:

• IR curves: GMM / static from ORE examples
• FX spot: GMM / static
• Equity spot: GMM / static
• Option vol surfaces: SSVI parameter dynamics / static from ORE examples
• Swaption/capfloor vol: GMM + SABR / static from ORE examples
• Credit / dividend / inflation vol: static only (gap types; no generation technique available)

Technique overrides per group allow a developer to test one asset class with a new technique while keeping all others static, without re-generating the full environment.

GMM parameters:

The control panel provides the following GMM parameters, configurable independently per asset class group:

• K: number of mixture components (integer, default 3; range 1–10)
• Training window: length in days (integer, default 1260 trading days = 5 years)
• Stationarity transform: levels / first-differences / log-returns (dropdown per group, with a sensible default: first-differences for rates, log-returns for spots and vols)
• PCA components: integer (shown only for high-dimensional groups — swaption cube, capfloor cube; default 10)

GJR-GARCH parameters:

The control panel provides direct parameter entry for the GJR-GARCH daily variance of each spot: ω (omega), α (alpha), γ (gamma), β (beta) and the starting variance. Each field accepts a floating-point value. Alongside the manual entry fields, the panel provides a "Calibrate from fixings" button per pair: clicking it runs the quasi-maximum-likelihood fit on five years of the pair's daily fixings and populates the fields with the calibrated values.

Tick frequency per asset class:

Each asset class group has an independent tick clock, configurable in the control panel. The tick clock determines how often that asset class generates a new data point during the scenario. The control panel provides:

• Tick mode: Fixed interval / Poisson / Regime-switching / Hawkes (dropdown per group)
• Fixed interval: interval field (seconds / minutes / hours / days, numeric)
• Poisson λ: mean ticks per time unit (float)
• Regime-switching: λ_active, λ_quiet, p_active→quiet, p_quiet→active (four floats)
• Hawkes: μ (base rate), α (excitation), β (decay) (three floats)
• Intraday profile: active window (e.g. 08:00–17:00 UTC) and quiet window outside those hours (time-of-day pickers)

Sensible defaults (matching typical market observation frequencies):

Asset class group	Default mode	Default interval
IR curves	Fixed	1 day
FX spot	Regime-switching	λ_active=12/h, λ_quiet=2/h
FX forwards	Derived — ticks when FX spot or IR curve ticks	—
Equity spot	Regime-switching	λ_active=12/h, λ_quiet=1/h
Equity forwards	Derived	—
Credit spreads	Fixed	1 day
Commodity forwards	Fixed	1 day
Inflation	Fixed	1 week
Option vol surfaces	Fixed	1 day
Swaption/capfloor vol	Fixed	1 day
Correlation	Derived	—

The intraday profile can be disabled (tick uniformly across 24 hours) for scenarios that do not require time-of-day fidelity.

Time range and scenario type:

• Start date / time: date+time picker (date only if no intraday profile active)
• End date / time: date+time picker (or start + duration)
• Scenario duration: convenience field (e.g. "1 week", "3 months", "1 year") auto-fills end date
• Scenario type: base (GMM free-draw) / stressed (pin high-vol component) / user-defined conditioning (see time evolution)
• Random seed: integer field for full reproducibility; "Random" checkbox for ad-hoc exploration (generates a seed from the system clock and displays it post-generation for repeatability)

Target sandbox:

The control panel provides a sandbox selector dropdown (see sandbox isolation) with an option to create a new sandbox inline. The currently selected sandbox name is displayed prominently alongside a badge showing its current state (empty / has data / compared).

Generate button:

The control panel provides a "Generate" button. Clicking it starts the generation pipeline in the dependency order from the implementation roadmap. The button transitions to "Stop" during generation. A progress area below the button shows per-asset-class status rows (e.g. "IR curves (EUR-ESTR): done", "FX spot (EUR/USD): running", "Equity vol surface (SP5): pending"). Each row shows a green checkmark, a spinner, or a red cross depending on status. A summary line shows elapsed time and estimated time remaining. On failure, the row shows the specific constraint or validation that was violated, with a link to the constraint in the quality dashboard (visualisation).

Generation writes into a Sandbox, an isolated, named market data environment that does not affect the live environment or any other user's session. This feature extends the sandbox concept with a synthetic market data generation owner.

The sandbox supports the following properties:

• A sandbox holds a complete set of market data for a contiguous date range.
• Multiple sandboxes can coexist; each is independently editable and independently deleteable.
• Generating into a sandbox is non-destructive and reversible. The previous contents of the target sandbox are replaced only when the generation completes successfully; a failed or aborted generation does not overwrite existing sandbox data.
• A sandbox can be compared against another sandbox or against the live environment. The comparison view shows divergence per quote key as a difference table and as overlay time series charts.
• A sandbox can be promoted to the live environment when the user is satisfied with the generated data. Promotion is gated: the quality dashboard (visualisation) must show green on all mandatory constraints before the "Promote" button is enabled.

The control panel's "Target sandbox" selector lets the user pick an existing sandbox or create a new one with a user-supplied name and optional description. This is the primary iteration mechanism: generate → inspect visualisations → adjust parameters → regenerate into the same or a new sandbox → compare sandboxes → promote.

ORE Studio provides visualisation surfaces to let users evaluate generated data before promoting it to production use. All visualisation surfaces read from the selected sandbox and update automatically when generation completes.

Time series viewer:

The time series viewer provides line charts of any selected quote type over the generated date range. It supports multi-series overlay (e.g. the EUR 5Y OIS swap rate across three generated sandboxes simultaneously). The date range and resampling frequency (daily / weekly) are configurable. A statistics panel alongside each series shows mean, standard deviation, min, max, skewness, and excess kurtosis computed from the generated series. The viewer accepts click-selection of any quote key from the ORE catalogue browser, and supports filtering by quote type prefix (IR_SWAP/RATE/EUR/, etc.).

Yield curve viewer:

The yield curve viewer shows a per-currency snapshot of the zero curve or par swap curve as a line chart with tenor on the x-axis and rate on the y-axis. It supports animated evolution through the generated date range using a forward-step button and a scrubber. Multi-sandbox overlay allows visual comparison of the same curve date under different generation parameter sets. Hovering over a point shows the underlying IR_SWAP/RATE quote key and its generated value.

Volatility surface viewer:

The volatility surface viewer provides a 3D chart with expiry on one axis, strike (or delta for FX surfaces) on the other, and implied vol on the Z-axis. It supports slice views: a fixed-expiry smile chart and a fixed-strike term structure chart. For swaption cubes, it provides a heatmap view with expiry on the x-axis, swap tenor on the y-axis, and ATM vol as the colour intensity. Animated evolution steps through generated dates. Cells where calendar-spread or butterfly violations are detected are highlighted in red on the 3D chart and in the heatmap. Violations are detected using QuantLib's BlackVarianceSurface::blackForwardVariance and SmileSection::density checks.

Cross-asset correlation heatmap:

The correlation heatmap shows the realised Pearson correlation matrix computed from the generated daily returns across all active asset classes. A reference panel alongside shows the same matrix computed from the historical training data. Cells are colour-coded from −1 (blue) to +1 (red). Clicking a cell opens a scatter plot of the two corresponding time series. This is the primary diagnostic for verifying that the joint GMM has produced the intended cross-asset correlation structure.

Quality dashboard:

The quality dashboard provides a per-asset-class checklist of all ten constraints. Each constraint row shows a green checkmark if the constraint is satisfied within tolerance, a red cross if violated, and a dash if not applicable for the current asset class selection. The constraint rows include a detail link that opens the relevant pair of time series or vol surface slice for investigation. A summary statistics panel shows, for each constraint, the maximum deviation found across the generated date range (e.g. "CIP maximum deviation: 0.3 pips" or "Real rate minimum: −0.8%"). The quality dashboard includes an "Export report" button that generates a PDF summary of the generated environment, including constraint check results, summary statistics per asset class, and sample yield curve and vol surface snapshots.

A future feature will provide a pre-defined portfolio of representative ORE trades (one per supported product type: vanilla IRS, cross-currency swap, FX forward, FX option, equity option, CDS, commodity forward, inflation swap, swaption). The generated market data will be used to price this portfolio. If all trades price without errors and produce economically plausible NPVs, the generated environment is declared validated. This is a separate piece of work and is explicitly out of scope for this sprint.

Scope: EUR OIS (EUR-ESTR) discount curve, USD OIS (USD-SOFR) discount curve, EUR/USD spot (FX/RATE/EUR/USD), EUR/USD FX forwards derived via CIP (FXFWD/RATE/EUR/USD at standard tenors), EUR swaption ATM normal vols (SWAPTION/RATE_NVOL/EUR/, selected expiry × tenor cells), USD swaption ATM normal vols.

Priced products: EUR vanilla IRS, USD vanilla IRS, EUR/USD FX forward, EUR/USD cross-currency basis swap (zero basis), EUR ATM swaption, USD ATM swaption.

Generation stack steps active: Steps 1 (IR), 3 (FX spot), 4 (FX forward derivation), 11 (ATM swaption vol, no smile, no frozen pool).

Exit criteria: A test run through ORE's Swaption and FxForward pricers produces non-NaN, economically plausible NPVs for all generated dates.

Scope: Phase 1 plus equity spot for S&P 500 proxy (EQUITY/PRICE/SP5/USD), flat dividend yield (EQUITY_DIVIDEND/RATE/SP5/USD/1Y ± noise), equity forward derived via carry (EQUITY_FWD/PRICE/SP5/USD), equity option vol surface via SSVI parameter dynamics (EQUITY_OPTION/RATE_LNVOL/SP5/USD/), EUR/USD FX option vol surface via SSVI parameter dynamics (FX_OPTION/RATE_LNVOL/EUR/USD/) quoted in the pair's conventions.

Priced products: European equity options, FX vanilla options, equity variance swaps, FX variance swaps.

Generation stack steps active: Phase 1 steps plus steps 5, 6, 7, 10.

Exit criteria: ORE EquityOption and FxOption pricers price without NaN. Put-call parity holds within 0.5 vol bp for ATM strikes across generated dates.

Scope: Phase 2 plus full EUR swaption cube (expiry × tenor × smile via SABR parameter GMM), EUR capfloor vol cube with computed shift, EUR EUHICPXT ZC inflation swap rates with seasonality, inflation YY rates derived from ZC.

Priced products: Bermudan swaptions (via ORE AMC), inflation-linked swaps, inflation cap/floors.

Generation stack steps active: Phase 2 steps plus steps 9 (inflation), 11 (full cube with smile). Frozen pool integration requires OD-3 resolved first.

Exit criteria: ORE InflationSwap and CapFloor pricers price without NaN. Real rate remains above −1% across all generated dates. Capfloor shift ≥ |min forward rate| at each generated date.

Scope: Phase 1 plus CDS spread curves for a sample reference entity (interim historical sampling method), static recovery rates, commodity forward curve for gold (COMMODITY_FWD/PRICE/GOLD/USD/ tenor strip) generated by GMM with carry consistency check.

Priced products: CDS, commodity forwards, commodity futures options.

Generation stack steps active: Steps 2, 8, 13. No hard dependency on Phases 2 or 3.

Exit criteria: ORE CreditDefaultSwap and CommodityForward pricers price without NaN. CDS hazard rate derived correctly from spread and recovery. Commodity carry check passes for gold.

Scope: Joint GMM across all active asset classes with block-diagonal or hierarchical covariance (resolve OD-1); realised correlation surfaces (CORRELATION/RATE) derived from joint draws; static base correlations for CDX.NA.IG.

Priced products: Basket equity products, CMS spread options, CDX/iTraxx tranches.

Generation stack steps active: Step 12 (correlation derivation); all prior steps using joint GMM architecture.

Exit criteria: Generated correlation matrix matches the historical correlation matrix within 10 percentage points for all major cross-asset pairs.