📘 About the Project

🧠 Inspiration

Prediction markets have quietly become one of the most trusted signals for forecasting real-world outcomes—from elections to sports to geopolitics. But in late 2024, something strange began happening: VCs and journalists started quoting AI prediction markets as if they were objective truth.

That raised a question that became the core of this project:

Do prediction markets forecast winners—or do they create them?

The idea that markets could shape the AI race—by influencing media coverage, VC funding, and talent allocation—was compelling, under-explored, and socially important. That became the spark.

🔨 How I Built It

The build had three major components:

1. Data Engineering

I ingested and cleaned a 59,000+ tick-by-tick dataset from Kalshi’s AI LLM markets:

prices (bid/ask/high/low)
volumes
open interest
timestamps

Nested JSON pricing required custom parsing:

df['close_price'] = df['PRICE'].apply(lambda x: (x.get('close') or 0) / 100)

2. Financial + Behavioral Modeling

I blended ideas from:

market microstructure
volatility modeling
herding behavior
Kelly criterion
market efficiency tests

with behavioral economics concepts like:

attention vs. conviction
speculative volatility
self-fulfilling feedback loops

For example, I defined an Influence Score:

[ I = V \times \sigma ]

where:

(V) = trading volume (public attention)
(\sigma) = volatility (uncertainty/speculation)

This captured narrative power, not just probability.

3. Machine Learning

Finally, I trained a Random Forest classifier to predict which AI lab would win. Surprisingly, the model favored Google (63%), despite OpenAI dominating headlines and volume.

📚 What I Learned

Three takeaways surprised me:

Markets are not mirrors — they are amplifiers. High volatility created attention cascades.
Prediction ≠ Influence. xAI had far lower odds than Google, yet a much higher influence score.
Financial markets and AI hype are entangled. Media → Markets → Capital → Products → Media formed a feedback loop that resembled:

[ x_{t+1} = f(x_t) ]

Self-fulfilling prophecy territory.

🚧 Challenges

This project broke my brain in several ways:

Data complexity: Tick-level pricing data is noisy, sparse, and timestamp-skewed.
Volatility extraction: Traditional finance metrics assume liquid assets; prediction markets are not liquid.
Causality vs correlation: Influence (attention) is easier to measure than power (action).
Modeling hype: There is no classical finance metric for “Elon tweeted and volatility exploded.”

The hardest conceptual challenge? Realizing that prediction markets don’t just reflect beliefs — they manufacture them.

🎯 Final Reflection

This project became more than a forecast. It became a lens into how economic belief systems shape technology, capital allocation, and ultimately the future of AI.

If prediction markets continue influencing AI funding decisions, then: