HypeInvest V2

A stock hype index that scores Reddit, YouTube, and news sentiment with Claude and blends them into one signal — built so any source failing degrades the index instead of breaking it.

Problem

Retail hype moves stocks before fundamentals do, but every sentiment source is a flaky third-party API — and a signal that dies whenever one of them does is worthless.

Built

A FastAPI backend with a pluggable scraper registry (Reddit, YouTube, news) running concurrently under per-source failure isolation, Claude scoring sentiment in batches, and a documented engagement-weighted index formula. Next.js 16 frontend.

Proof

Every API response reports sources_used and sources_failed, so degradation is visible instead of silent. Rebuilt from the HackUTD VIII project that won Goldman Sachs' Investment Strategy challenge.

Context

The original HypeInvest came out of HackUTD VIII, where it won Goldman Sachs' Investment Strategy challenge: follow retail hype as a leading signal. V2 is the production-shaped rebuild of that idea, and its real subject is a less glamorous question — how do you blend several unreliable data sources into one number without the weakest source setting your uptime?

For any ticker Finnhub can resolve, the index answers: how loudly is the crowd talking, and how do they feel? It's a gauge from −100 to +100, deliberately not a trading system — the docs warn against using hype in isolation.

What I built

Scrapers implement a common base class and live in a registry — Reddit via asyncpraw, YouTube via the Data API, and news via NewsAPI — pulling up to 125 items per query from the trailing seven days. They run concurrently, and one scraper throwing doesn't touch the others: failures are isolated per-source and surfaced in the response as sources_used and sources_failed.

Claude Haiku scores sentiment in batches of twenty texts per call, with output clamping, padding, and a parse-failure fallback; if no API key is configured the pipeline degrades to neutral sentiment instead of erroring. The index formula is documented in the code — sign(perception) × tanh(4 × popularity × |perception|) × 100 — weighting engagement per platform, with Finnhub as the primary stock-data source and yfinance as fallback. Pydantic v2 schemas type every boundary.

Decisions

• Degrade, don't break

  Scrapers run under isolation so an exception in one becomes a smaller sample, not a 500. The failure shows up in the response payload — visible degradation over silent failure or a dead endpoint.

• A registry, not a hardcoded pipeline

  Adding a sentiment source is one class implementing the scraper base. The blend logic never changes when the source list does.

• Batch the model, bound the cost

  Twenty texts per Claude call with clamped, padded JSON-array output and a fallback when parsing fails. Per-item calls would multiply cost twentyfold for the same signal.

• Show the math

  The index formula sits in the code with its reasoning — tanh keeps runaway popularity from saturating the scale, engagement weights keep one viral post from being the whole signal. A number you can't derive is a number you can't trust.

• A gauge, not a trading system

  The output is deliberately framed as one input among many. Overclaiming a sentiment score into trade advice is how fintech demos become liabilities.

What this proves

Signal engineering is mostly failure engineering: isolate unreliable sources, make degradation visible in the API contract, bound LLM cost by design, and document the formula so the number can be audited.