xcadence

A continuously-running simulated exchange in music royalty shares, built to make a specific argument: that the standard toolkit for valuing an asset class with heavy-tailed returns quietly misleads you, and that you can see it happen if the market is real enough to watch.

The market runs on its own. Prices move whether or not anyone is looking, a server-side clock advances simulated time at up to a month per minute, thirty-two synthetic participants trade continuously, and the whole state survives a restart.

Stack — Next.js 16 (App Router) · TypeScript · Tailwind · Recharts · SQLite via Prisma. No other runtime dependencies: the seeded RNG, log-normal and Pareto sampling, LMSR market maker, Kaplan-Meier estimator with Greenwood intervals, IRR solver and mean-variance optimiser are all implemented directly.

---

What it does

A market that runs itself. A setInterval inside a long-lived route handler advances simulated time at 1×, 60×, 1440× or 43200×. Every tick moves artist fundamentals, runs synthetic order flow and accrues royalties on schedule. State write-behinds to SQLite, so killing the server and restarting resumes exactly where it stopped. Server-Sent Events push live quotes to the browser.

An AMM instead of an order book. Each artist has a Logarithmic Market Scoring Rule maker: price(q) = vMax · σ(q/b), C(q) = vMax · b · ln(1 + e^(q/b)). An order book with no resting orders quotes nothing at all, which on four hundred thin listings means an empty screen on exactly the names the venue exists for. LMSR always quotes both sides at a bounded, known subsidy — the cost is that your own size moves the price, and the ticket shows you the exact slippage before you commit.

Heavy-tailed returns, on purpose. Monthly log-growth is Δlog L = μ + κ·(log L* − log L)/12 + σ·Z + 1{breakout}·log(Pareto). Across a run the mean listener count sits roughly 30× the median. Reversion is asymmetric — 0.35 pulling up toward an artist's quality level, 2.0 pulling down from above it — because with symmetric reversion the multiplicative jumps stack and the top of the universe runs away past any listener count a real artist has had.

Hidden ground truth. Every artist carries a trueQuality and a real hazardRate that nothing in the normal interface can read. The bots and every valuation see only observable fundamentals and a hazard rate inferred from tier. That gap is the point: it is why the market can stay wrong, and why the run inspector has something to measure.

The quantitative work

Tool	What it does
DCF	Royalty stream discounted with a decaying growth rate and a survival factor. Sensitivity sweep marks where valuation halves.
Monte Carlo	10,000 paths, fan chart, terminal distribution binned at P99. Reports the mean−median gap explicitly.
Survival	Kaplan-Meier on recorded exits, right-censored at current age, segmented by tier at debut.
Survivorship bias	The same cohort computed twice — survivors only vs everyone who was there.
Diversification	Correlation matrix and portfolio variance against holding count, falling to a market-risk floor rather than to zero.
Adverse selection	Sliders for seller information and offer price; shows the accepting pool degrade and the market unravel.
Portfolio	Mean-variance over the live book: alpha, beta, appraisal ratio, and a risk/return scatter.

Findings the simulation actually produces

• Survivorship bias overstates mean return by ~104%. Survivors-only mean +57% against a full-cohort mean of −23%. Both from the same recorded run; exits are written with a date and never deleted.
• Royalty offerings: 56% return less than invested, median −51% against a mean of +262%, with a single 21× outcome carrying 65% of cohort value. Nothing is weighted to produce this — offerings are priced off a tier-average hazard rate while each artist's real hazard is drawn individually, so sellers who are worse than they look are the ones happy to sell.
• Information is rewarded. Fundamental bots returned +504% against momentum at −153% and noise at −77%.
• The market underpays for quality it cannot see. Hazard error is positive in every quality decile and shrinks monotonically with quality, from +1.3% a month at the bottom to +0.02% at the top.

Why Black-Scholes is not used

Three independent reasons, any one disqualifying: there is no traded underlying (you cannot buy a royalty stream, only claims on it — and the claim is the instrument); there is no replicating portfolio (nothing here can be hedged, and exit risk is not spanned by any tradeable asset); and returns are not log-normal (power-law jumps plus a hazard of absorption at zero). The methodology page covers this and the limitations of every model used instead.

---

Running it

npm install
cp .env.example .env
npx prisma migrate dev
npm run dev

Then http://localhost:3000. The engine boots with the server, seeds a universe on first run, and starts ticking.

Two rosters

ROSTER_MODE in .env selects what the exchange lists.

• real (default) — every established rap, R&B and house artist (2,000+), with photographs and biographies from Wikipedia, catalogues from MusicBrainz and sleeves from the Cover Art Archive. No API keys required. A background pass fills the cache in a couple of hours at MusicBrainz's one-request-per-second limit.
• demo — generated names, no real people, and no outbound requests at all. The market, the engine and every quantitative tool behave identically; only the names and the profile lookups change. Demo mode also restores continuous artist entry, which the real roster cannot have — adding a listing there would mean inventing a name.

Every figure in the application is simulated in both modes. In real mode the biography, photograph and discography are genuine and everything else — listener counts, royalties, prices, tier, volatility, and events such as dropped by label — is output from a seeded random number generator and describes nobody. Artist pages separate the two under their own headings.

Deployment

docker build -t xcadence .
docker run -p 3000:3000 -v xcadence-data:/data xcadence

The image defaults to ROSTER_MODE=real. On first boot it works through the roster at MusicBrainz's one-request-per-second limit — roughly nine minutes — so photographs and discographies fill in progressively while the market runs normally. The cache lives on the mounted volume, so redeploys against the same volume start warm.

Every page carries a disclosure naming which fields are genuine and stating that all market data is generated. Set ROSTER_MODE=demo if you would rather publish generated names instead.

This will not run on a serverless platform. The clock is a setInterval in a long-lived process and the state is a SQLite file, so it needs a container that stays warm and a writable volume — Railway, Render, Fly.io or a VPS. Functions that freeze between requests stop the clock; ephemeral filesystems lose the run.

---

Design notes

A few decisions that were not obvious, recorded because the reasoning is the interesting part:

• Contracts carry a per-artist unitScale fixed at listing. Without it a superstar contract prices near 90,000 credits and a dormant one at 0.02, and a single account cannot trade both.
• Bots rank opportunities on |ln(fair/price)|, not fair/price − 1. The latter is unbounded above but floored at −1 below, so every overvalued artist sorted to the bottom of the shortlist and nothing was ever sold.
• Bot position limits size off equity, not cash. Sizing off cash froze the fundamental bots permanently once they were fully invested — including out of selling names they did not hold.
• A failed flush is retried, not discarded. The whole flush is one transaction, so a failure rolled back lastMonthKey while the world kept advancing; the next reload replayed months already written. World and database had drifted to 218 vs 393 active artists before this was caught.
• Artists share a common factor. Without an industry-wide shock and aggregate passive flow, every pairwise correlation sits at zero and the diversification tool shows variance falling to nothing, which is not what diversification does.

Licence and attribution

Artist biographies and photographs come from Wikipedia (CC BY-SA), catalogue data from MusicBrainz (CC0) and cover art from the Cover Art Archive. Artist pages link back to the source for each. No affiliation with, or endorsement by, any artist named in real mode is implied.

Virtual currency sandbox. Not a financial product.