Trading Screen
Login Screen

Inspiration

We love music and one thing that we always argued about is when someone started listening to an artist. For example, if an underground artist suddenly started having hundreds of thousands of listeners, one of us would say we started listening to them when they only had 40 thousand. Taking inspiration from the stock market and the Bloomberg Terminal, we wanted to create a system that not only showed this metric but could also show support for artists as well.

What it does

Early-Signal.FM is a simulated stock market platform for upcoming and established artists. Each player starts off with 10,000 credits, and they're allowed to invest those credits into their favorite artists. By taking in news such as album drops, collaborations, and drama, fans can see how an artists develops their reputation over time. There is also a foresight section that predicts a higher and lower ground on how an artist would perform the next day.

You can also compare your own portfolio to other fans through the leaderboards. There is one for overall net worth, and one for when fans started listening to a specific artist. Additionally, there are bots that use different investing strategies, so try to beat them as well!

How we built it

A Spring Boot API and PostgreSQL run the market, and a React trading terminal is the frontend. A Python ML service runs alongside: after every game day it estimates each artist’s growth from listener data, fits a Gaussian process to produce a 7-day forecast cone, then lets four bots (including our AI trader) trade through the same API players use. Everything ships in one Docker container on Render.

Challenges we ran into

We planned to train on real streaming data, but Spotify’s API no longer offers listener history, follower counts or popularity, and public datasets only cover chart-topping hits, not underground artists. So we built a realistic simulated market instead, with real artist names and simulated numbers, and designed the model to plug into a real data feed later. The simulated market is mostly noise, so our early forecasts couldn’t beat “the price stays flat,” and our first AI trader lost to simple buy-and-hold by overtrading into fees. Backtesting helped us fix both.

Accomplishments that we're proud of

An AI trader that earned the most of any strategy across 80 simulated seasons (+44% vs. +32% for buy-and-hold), and a forecast that ranks artists by true growth twice as well as using prices. It’s live and running in real time on a free-tier server.

What we learned

The data you want often isn’t available, so build models that can swap in a real source later. A single season is mostly luck, so we judged strategies over 80 simulated seasons which exposed our AI’s overtrading and turned it into the top earner.

What's next for Early-Signal.FM

Being able to actually implement this with real artist data by collecting their monthly listener count over time.

Built With

Submitted to

Cal Hacks 13.0

Created by

I led the build by defining the API contract early, which let all three of us work in parallel. Using Spring Boot 3 and Java 21, I designed the PostgreSQL schema and the REST API behind user portfolios, artist listings, instant buy/sell trades with a 1% fee and a 30% position cap, a daily price engine driven by listener growth and volatility, events, and the leaderboard. I also set up the forecasts table that stores our Gaussian Process Regression predictions for the chart forecast cones, plus an admin "Advance Day" endpoint so we could simulate market days live during the demo.

Joshua Hyejun Park
Joseph Yojune Jung
Matthew Bersalona