Help us double down on what's working, instead of guessing. Takes 5 seconds, totally optional.
Build vs Buy: What It Actually Costs to Build Dealer-Positioning Infrastructure
What a fund takes on building gamma exposure and volatility-surface infrastructure in-house: the seven components, the maintenance that never ends, and when building is genuinely right.
The build case is seductive because the maths is public. Gamma exposure is open interest times gamma times a multiplier, summed by strike. You can write that in an afternoon against a broker chain endpoint and get a plausible-looking chart the same day.
The chart is not the product. The product is the same number, computed identically, every minute, for years of history, that you are willing to put a position on. That is a different engineering problem, and it is the one this article prices.
The seven components
1. Chain ingestion
You need every listed contract for every symbol you care about, continuously. That means an OPRA-derived feed or a vendor API, contract lifecycle handling as strikes are added and expiries roll, and a store that can serve a full chain fast enough to compute on. The naive version polls a broker endpoint and works until you add the twentieth symbol or the first 0DTE session.
Ongoing cost: feed licensing, plus the operational reality that this component fails on the days it matters most - high-volume sessions when chains are largest.
2. Implied volatility solving
Every contract needs an implied volatility before it can have a Greek. That is a root-find per contract, per snapshot, across thousands of contracts. The mechanics are textbook; the engineering is not. You need a bracketing fallback when Newton-Raphson fails, a decision about which price to solve against, a policy for one-sided and crossed markets, and a rule for contracts that admit no solution. Get the last one wrong and a handful of broken quotes quietly distort an aggregate.
3. Greeks
First-order Greeks are straightforward. Second and third order are where conventions start to matter: per-day or annualised theta and charm, vega per point or per unit, sign conventions for vanna. These choices are individually trivial and collectively the reason two systems disagree by a factor you will spend a day chasing.
4. Volatility surface fitting
If you want a smoothed surface rather than raw per-contract implied volatilities, you need a parameterisation, a per-expiry fit, a forward recovered from put-call parity, wing constraints to keep the fit sane where quotes are thin, and a decision about whether to enforce butterfly and calendar arbitrage checks. This is the single component most teams underestimate.
5. Exposure aggregation
The actual GEX/DEX/VEX/CHEX layer: per-strike aggregation, a stated dealer-positioning convention, level derivation (gamma flip, call and put walls), and the filtering policy for thin strikes. The arithmetic is easy. The conventions are a methodology you now own and must defend internally when a PM asks why your number differs from a screenshot they saw.
6. History
This is the component that turns a project into a programme. Live analytics answer "what is dealer positioning now". Research needs "what was it at 14:32 on a Tuesday in 2019", which means storing snapshots at resolution, for years, across your universe, and being able to query any timestamp fast. It means a time-series store, a backfill campaign, and a decision about what to keep: full chains are enormous, aggregates alone foreclose questions you have not thought of yet.
It also means confronting point-in-time integrity honestly: restatement policy, survivorship, and keeping derived statistics such as percentiles free of look-ahead. Most in-house archives get this wrong quietly, because nothing fails - the backtest just flatters you.
7. Data quality
The component nobody scopes. Feeds go stale silently. Open interest arrives late or not at all. A holiday session produces a chain that looks fine and is wrong. Without reconciliation against an independent source, you will not discover any of it until a strategy behaves strangely and someone spends two days proving the data was bad. Building this properly means a second source to compare against, thresholds, alerting, and the judgement to treat "comparison source returned nothing" as a blind spot rather than as agreement.
Where the estimate goes wrong
Teams scope components 1 through 5 and deliver them, roughly on time. Components 6 and 7 - history and data quality - are what convert a working prototype into something a strategy can be run on, and they are typically discovered rather than planned. They are also the two that never finish, because they are operational rather than constructive.
The part that does not end
A build is not a project with a completion date; it is a permanent line item.
Feed and vendor licensing, renegotiated as coverage grows.
Operational ownership. Someone is on the hook when the chain does not arrive before the open. That someone is a quant developer whose time you were spending on research.
Corporate actions and market structure. Splits, special dividends, new expiry cycles, symbol changes, and the day a new product like a shorter-dated expiry appears and your bucketing assumptions quietly stop holding.
Storage growth. Minute-resolution options history compounds. The archive that was comfortable in year one is a budget conversation in year three.
Institutional knowledge risk. The person who chose the conventions leaves, and the methodology becomes folklore.
When building is genuinely right
There are real cases. A vendor pretending otherwise is selling, not advising.
Build when
The methodology is the edge. If your dealer-positioning model differs deliberately from convention and that difference is the alpha, you cannot outsource it. Buying someone else's convention would erase the thing you are trying to express.
You need per-contract control. If your strategy operates on individual contracts with bespoke filtering rather than aggregates, an analytics layer is the wrong shape and you want raw data plus your own stack.
Latency is the strategy. Below the millisecond regime, any external dependency is the bottleneck. Colocate and build.
Buy when
The analytics are an input, not the edge. If dealer positioning is one feature among many in a model, building the pipeline is undifferentiated work.
You need history immediately. Backfilling years of minute-resolution options history is measured in months and machines, and it cannot be compressed by hiring.
The team is small. One or two quant developers spending a quarter on infrastructure is a quarter not spent on strategy, and the infrastructure still needs an owner afterwards.
You want the disagreement resolved. Comparing your build against an external source is itself valuable; two independent computations that agree is a stronger position than one you cannot check.
The hybrid most desks actually land on
The common outcome is not build-or-buy but both. Buy the computed layer to start research immediately and to have an independent number to reconcile against; build selectively where you have a genuine methodological view. That sequencing matters: teams that build first and evaluate vendors later usually discover the vendor after they have already paid the construction cost, and then keep the build alive out of sunk-cost rather than merit.
A cheap first test: before scoping a build, take a vendor's history for one symbol and one year, and run the study you actually care about. If the study answers your question, the build was going to be infrastructure for a result you could have had in a week. If the study cannot be expressed with the vendor's conventions, you have learned precisely which component you need to own.
What FlashAlpha replaces, specifically
To be concrete rather than promotional: FlashAlpha is components 2 through 7. Chain ingestion is upstream of us; the computed layer, the surface fit, the exposure aggregates, the history, and the reconciliation are what a subscription is.
Computed analytics - GEX, DEX, VEX, CHEX by strike, gamma flip, call and put walls, max pain, 0DTE regime, SVI surfaces, variance risk premium, and 15 BSM Greeks across 6,000+ US equities and ETFs plus CME index futures.
History - point-in-time replay from 2017 at minute resolution on the flagship symbols, with response shapes identical to live, so a validated strategy ships without re-derivation.
Data quality - cross-source reconciliation through the session and endpoints that fail closed rather than serve a stale value, documented in the methodology whitepaper.
The conventions, written down - the whitepaper states the dealer-positioning assumption and the model's limitations, so you are adopting a documented methodology rather than inheriting an undocumented one.
What it does not replace: a genuine methodological edge, per-contract execution logic, or a sub-millisecond path. Those are yours to build, and should be.
The build decision is rarely about whether the team is capable - it usually is. It is about whether dealer-positioning infrastructure is the thing you want that capability pointed at for the next two quarters, and whether you want to own its maintenance permanently afterwards. If the analytics are an input to your edge rather than the edge itself, the cheapest possible test is to run your actual study against someone else's history first. See the institutional datasheet for coverage and delivery, or the methodology whitepaper if you want to evaluate the conventions before the data.
