Metafide was founded in 2023. Before writing a line of product code, Marcus and I spent several months mapping the problem space we had both hit repeatedly during years of building quantitative risk systems for institutional desks. The frustration we kept returning to was straightforward to state but difficult to solve: vol research was always fragmented. The equity surface lived in one place, rates vol in another, FX in a third. Cross-asset correlation data was either stale, proprietary and inaccessible, or both. And the tools available for connecting these pieces into a unified morning view simply did not exist in a form useful to a vol desk.
This post is about what we learned building the research framework from nothing. It is not a product roadmap. It is an account of the concrete problems we hit, the architectural decisions we made, and where we think the hard problems still are. Helena joined us early to lead the research methodology side, and a significant portion of what follows reflects her work on signal design and surface calibration.
The Data Problem Is Worse Than the Modeling Problem
The assumption going in was that the hard work would be in the models: surface interpolation, regime detection, cross-asset signal integration. We were wrong. The data problems consumed more engineering time and introduced more subtle errors than the modeling choices.
Options chain data is messy in ways that are easy to underestimate until you have worked with it at scale. Raw options data includes stale quotes that were not updated before close, crossed markets during illiquid windows, dividend and corporate action adjustments that affect put-call parity relationships, and exchange-specific identifier conventions that do not map cleanly across asset classes. Before any vol surface can be constructed, the underlying options chain data needs cleaning, normalization, and a set of quality filters that identify and exclude or correct anomalous quotes.
For equities this is tractable, if tedious. The major exchanges have good coverage, and the data ecosystem around US equity options is well-developed. For rates options (swaptions, caps, floors), the data is fundamentally different: it comes primarily from OTC markets where quotes are indicative rather than executable, and the swaption cube requires interpolation across dimensions (expiry, tenor, moneyness) where data points are sparse. A swaption vol surface for a 2-year expiry on a 10-year swap is not observed at every strike. It is a combination of the broker quotes you can get and the interpolation model you choose to fill the gaps, and the two are not separable once you have a number.
FX options add another layer: the quoting convention for FX options is delta-space, not strike-space, and the conversion between the two requires knowing the forward price and domestic and foreign interest rates at each expiry. Small errors in the input rates produce material errors in the converted strikes for long-dated FX options. We rebuilt the FX surface construction pipeline twice before we were satisfied with the conversion accuracy.
Architecture Principles We Arrived At
Several architectural principles emerged from our first year of building that we now treat as non-negotiable for the research framework.
The first is data provenance tracking. Every number that flows into a forecast must carry metadata about its source, timestamp, and quality flag. In a multi-asset framework where data arrives from different sources at different times with different update cadences, a forecast that mixes observations from different time windows without flagging the inconsistency will produce silently wrong results. We learned this from a specific incident where a stale FX vol surface (from an earlier timestamp because the data feed had a gap) mixed with a current equity surface in a cross-asset correlation signal, producing a signal that looked right numerically but was internally inconsistent by several hours. The provenance metadata layer now prevents this class of error by design.
The second principle is clean separation between the data layer, the signal layer, and the forecast layer. Early in the build, these were entangled: surface construction code called signal functions directly, and forecast code reached back into the data store for raw inputs. This made debugging and validation extremely difficult because a change in data handling could propagate to forecast outputs through multiple indirect paths. We refactored to enforce strict interface boundaries: the data layer produces normalized, quality-flagged data objects; the signal layer consumes data objects and produces signal objects; the forecast layer consumes signal objects. Nothing crosses these boundaries in reverse.
The third principle is that every pipeline stage must produce outputs that can be examined independently. A vol surface should be inspectable on its own. A signal score should be interpretable in isolation. A forecast should be explainable from its inputs. When the end-to-end result looks wrong, you need to be able to isolate which stage introduced the error. Without independent inspectability at each stage, debugging a multi-asset pipeline becomes an exercise in frustration.
Cross-Asset Coverage: Scope Decisions We Made Early
A multi-asset vol framework can expand indefinitely in scope. Every asset class and every market segment has its own vol characteristics, its own data complexity, and its own set of practitioners who need to understand it. We made an early decision to go deep on five asset classes rather than thin across everything: equities (single-stock and index), rates (swaption cube and Treasury options), G10 FX, energy and metals commodities, and credit (IG and HY CDX vol). These five cover the major risk positions that institutional desks are managing, and they have enough cross-asset interaction to justify joint modeling.
The interaction between rates and equity vol is particularly rich. The swaption vol cube and equity index vol surfaces share macro sensitivity and react to the same set of central bank policy signals, but through different transmission mechanisms. Equities respond through the discount rate channel and risk appetite. Rates vol responds through the policy expectations channel and term premium uncertainty. A framework that monitors both simultaneously picks up more signal about the macro vol environment than one that monitors them separately.
We explicitly excluded crypto and digital assets from scope. The vol dynamics in crypto markets are structurally different from traditional financial markets (24-hour trading, different microstructure, regulatory uncertainty that creates discontinuous vol behavior), and mixing them into the same framework would require a separate modeling track that would dilute rather than enrich the institutional-desk focus.
The Signal Design Process
Helena's approach to signal design started from a practitioner observation rather than a statistical screen. The question was not "which variables have the highest historical correlation with future vol?" but rather "what information does a senior vol desk strategist actually use in their morning assessment, and which of those inputs can be made systematic and continuous rather than judgment-based?"
That framing produced a different candidate signal set than a pure statistical screen would have. The desk strategist looks at: how has the vol term structure moved in the last 48 hours? What is the options flow positioning telling us about near-term directional sentiment? What are the macro surprises in the economic release calendar doing to consensus expectations? And what are the cross-asset signals saying about correlation regime stability?
Converting these qualitative observations into systematic signals required both methodological work (how do you measure options flow positioning cleanly? how do you score macro surprise relative to current consensus?) and calibration work (what is the appropriate weighting of each signal relative to realized and implied vol in the final surface forecast?). The signal design and the surface forecast architecture are not independent: the signals were designed with the forecast integration in mind, and the calibration weights were chosen with the signals' information structure in mind.
What We Have Not Solved and Are Honest About
Vol forecasting at a point-in-time level for short horizons (1 to 5 days) remains substantially noisy. The surface-level forecast we produce is more useful for directional risk assessment (is vol likely to be elevated or compressed relative to current levels over the next week?) than for precision calibration at a specific strike and expiry. We communicate this to platform users directly. The research output is not a precision instrument; it is a structured, updated view of the vol environment that the desk uses as one input among several.
Regime transition forecasting remains the hardest problem. Detecting that a regime change has happened is tractable. Forecasting that a regime change will happen in the next 5 to 10 days is substantially harder. The leading indicators we have identified improve the probability of early detection, but they do not eliminate the possibility of missing a rapid transition. The platform communicates regime uncertainty as a range, not a point, and flags the confidence level of the current regime classification explicitly.
We started Metafide to fill a gap in the research tools available to institutional desks, and we are early in building toward that goal. The framework we have built reflects what a small team with direct experience on both the buy-side and sell-side thought was most valuable to get right first. The research published here is part of how we think in public about the problems we are working on. Feedback from institutional researchers who are working on the same problems is the input we find most valuable.
This article is research analysis only and does not constitute investment advice. Metafide does not manage money or execute trades.