Crypto Options Analytics Platform

Deribit is the dominant venue for crypto options, but its raw feed gives you no way to reason about the surface — where volatility is rich or cheap, how a multi-leg structure pays off, or whether a given contract is trading away from its theoretical value.

This platform turns that feed into a live analytics desk. It runs in production as four containers behind nginx.

What it does

• Subscribes to ~1,300 BTC and ETH option contracts simultaneously over Deribit’s WebSocket API
• Processes 200–400 market updates per second through a bounded Go worker pool
• Computes a Black-Scholes theoretical price for every contract on the hot path and flags its deviation from the exchange mark
• Maintains a live implied-volatility surface across the full strike × expiry grid, rendered as an interactive 3D plot
• Lets you assemble a multi-leg position and watch its payoff diagram and breakeven points update in real time
• Persists every tick to TimescaleDB with continuous aggregates at 1 min / 15 min / 1 hour

Engineering

Every optimisation below came out of a pprof profile rather than a guess.

The lookup table is my favourite tradeoff in the project. math.Erf was the hot spot, so I precomputed the standard normal CDF across the range -5 to +5 at 0.001 resolution — an 80 KB array, small enough to stay cache-resident, roughly 6× faster. Naive table lookup costs accuracy, so interpolating between adjacent entries recovers an order of magnitude (0.05% → 0.005% error) for about 5 ns. Space, time, and accuracy all quantified.

I also had to be honest about what “latency” actually meant. The metrics dashboard reported 1.1 ms per message while pprof said 95 µs p99. Both numbers were real — the dashboard was including queue wait, which turned out to be 84% of the total.

Queue wait dominating is healthy for an async worker pool (workers idle-waiting, not backlogged), but reporting it as processing time is misleading. I instrumented each stage separately so the number could actually be defended.

Beyond that: a three-tier cache (in-process LRU → Redis with Pub/Sub fan-out → TimescaleDB hypertables), an OptionState struct laid out by access frequency so a filter sweep over 1,300 contracts touches minimal cache lines, and a 1,300-row live table held at 60 fps through windowed virtualization.

The 7-day z-score lookback

Raw price history is hard to read. A $200 move in BTC means nothing on its own — it matters only relative to how that contract normally behaves. So every historical endpoint computes a z-score over a 7-day lookback window: each time bucket is normalised against the whole window’s mean and standard deviation.

Concretely, for each metric — mark price, IV, delta, open interest, theoretical price, price deviation — the SQL buckets ticks with time_bucket, computes the window mean and STDDEV_SAMP, then turns every point into (value − mean) / stddev. The result is unitless, so a +2 on an illiquid far-dated contract and a +2 on BTC’s front month are directly comparable: both are “two standard deviations from their own baseline”.

The frontend renders those bands as a chart rather than a number, with reference lines at the mean and ±2σ and colour-coded interpretation ranges.

The point of the feature isn’t the formula — it’s that a trader comparing a dozen strikes needs a single normalised axis, and the 7-day window is long enough to smooth intraday noise while short enough to stay responsive to the current regime.

The quantitative work

Black-Scholes is the live pricer. The deeper modelling work targets the Bates model — Heston stochastic volatility plus Merton Poisson jumps — priced with the Fourier-cosine (COS) method of Fang & Oosterlee (2008). COS expands the risk-neutral density in a cosine series, collapsing European pricing to a single sum whose truncation range and payoff coefficients are computed once and reused across the entire strike ladder — dramatically cheaper than Monte Carlo.

That implementation includes the “little Heston trap” branch correction for the complex logarithm (the naive branch makes long-dated prices oscillate and diverge), cumulant-based adaptive truncation widened for crypto’s fat tails, adaptive Fourier term counts for short-dated options, and Feller condition validation to keep the variance process strictly positive.

The genuinely interesting part is regime. Equity-calibrated parameters price BTC options 3–5× below market: crypto volatility runs roughly 10× equity’s, mean-reverts about 3× slower, shows almost none of the leverage effect that anchors equity models, and jumps around 3× as often.

Status

Being able to name the exact bug seemed more useful than a green checkmark.