About
MarketPulse is a real-time policy intelligence and prediction-market platform that listens to policymakers, identifies unexpected changes in their policy stance, and connects those signals to prediction markets.
Inspiration
When the Fed Chair speaks, prediction markets on the next interest-rate decision can reprice within seconds. Generic sentiment analysis misses an important part of the context: the same statement can mean very different things depending on who says it. “Inflation is too high” may be completely normal for a historically hawkish speaker but a meaningful shift for a historically dovish speaker.
MarketPulse was built around the idea that the most useful signal is not simply what a policymaker says, but how unusual that statement is for that specific policymaker.
What it does
MarketPulse continuously turns live policy speeches into measurable, actionable signals.
The system can listen to a Fed livestream, YouTube video, uploaded recording, saved transcript, or live microphone input. As the speaker talks, ElevenLabs Scribe v2 Realtime transcribes the speech and passes the transcript to Gemini for analysis.
Gemini evaluates each passage on a -1 to +1 policy-stancе scale, ranging from dovish to hawkish. It uses surrounding passages for context and distinguishes the policymaker's comments from reporters' questions. MarketPulse can also incorporate current FRED macroeconomic data, including inflation, unemployment, interest rates, payrolls, and Treasury yields, so statements are evaluated in the context of the current economic environment.
The key part of MarketPulse is its speaker-specific baseline. Instead of comparing every statement against a generic definition of hawkish or dovish language, the system learns what an individual speaker normally sounds like from their previous speeches and press conferences. Each new stance score is converted into a z-score relative to that speaker's historical mean and standard deviation. This allows MarketPulse to distinguish between a predictable statement and a genuine departure from a speaker's normal position.
When a statement reaches the project's surprise threshold of |z| ≥ 2, MarketPulse flags it as a potential policy surprise. For example, a normally dovish speaker making an unusually hawkish statement can generate a positive surprise even if the statement itself would not look particularly hawkish to a generic sentiment model.
MarketPulse then connects the surprise to prediction markets on Kalshi and Polymarket, focusing on markets related to upcoming Federal Reserve decisions. Markets can be selected before a speech begins, allowing the system to react without first searching through thousands of available contracts.
The system supports dry-run, demo, and live trading modes. When trading is enabled, MarketPulse applies safeguards including per-order and daily limits, slippage limits, duplicate-trade protection, an opposite-direction confirmation rule, and a kill switch.
MarketPulse also looks beyond the speaker. It combines the policy signal with macroeconomic conditions from FRED and crowd sentiment from Bluesky and Reddit, giving the user a broader view of how a statement fits into the current economic and information environment.
Everything is displayed through a live web dashboard. Users can see the live transcript, current speaker stance, historical baseline, z-score, economic context, crowd sentiment, relevant prediction markets, and orders as they happen. The system also supports replaying previous speeches and press conferences to analyze how MarketPulse would have responded in real time.
Every signal and order is stored for later analysis. Tiger Data stores the time-series history, Backboard provides searchable memory through the “Ask the Desk” feature, and Solana devnet can be used to create verifiable receipts for signals.
In short, MarketPulse turns a policymaker's live speech into a structured pipeline:
Speech → Transcription → Policy Stance → Speaker Baseline → Surprise Detection → Market Identification → Execution → Historical Record
How we built it
MarketPulse is built primarily in Python, with a web dashboard hosted on Hostinger. It combines ElevenLabs for transcription, Gemini for policy analysis, FRED for macroeconomic data, Bluesky and Reddit for crowd sentiment, Kalshi and Polymarket for prediction markets, Tiger Data for historical storage, Backboard for memory, and Solana for verifiable receipts.
The system was designed so that transcription, analysis, market selection, trading, and data storage can operate together in real time while the dashboard continuously publishes the current state.
Challenges we ran into
The main challenge was accurately interpreting short statements without losing their surrounding context. We also had to make market selection reliable enough to identify the specific upcoming FOMC contracts relevant to a surprise.
What we learned
The project showed us that context and speaker-specific history can be more important than generic sentiment when trying to identify meaningful policy surprises.
What's next for MarketPulse
Future versions could expand speaker baselines to more Fed officials and Treasury officials, track the historical accuracy of every signal against actual policy decisions, improve market recommendations, and move the live worker entirely to hosted infrastructure.
Log in or sign up for Devpost to join the conversation.