What This Project Does

A plain-language explainer

The one-sentence version

I built a system to figure out whether you can make money by betting on crypto prices in a prediction market, using math instead of speed — and I'm being very careful to make sure the answer is real before putting any actual money on the line.

What's a prediction market?

A prediction market is like a stock exchange, but instead of buying shares of a company, you buy contracts that pay out based on whether something happens. On Polymarket (the platform I study), you can buy a contract that says "Bitcoin will finish this 4-hour window higher than where it started." If you're right, you get $1. If you're wrong, you get nothing.

The price of that contract is basically what the crowd thinks the probability is. If the contract trades at $0.85, the market is saying "there's roughly an 85% chance this happens."

What's the "favorite-longshot bias"?

This is the single most well-documented pattern in betting markets, going back decades:

• Favorites (outcomes the crowd thinks are very likely) are slightly underpriced — they win more often than their price implies
• Longshots (outcomes the crowd thinks are unlikely) are overpriced — they win less often than their price implies

Why? People love the lottery ticket. Paying 5 cents for a shot at $1 feels exciting even when the math says you'll lose. Meanwhile, paying 95 cents to win 5 cents more feels boring, so fewer people do it — which means that 95-cent contract is actually a decent deal.

This isn't a theory or a hunch. A 2026 study of 300,000+ contracts on Kalshi (a similar platform) confirmed it with real data: longshot buyers lost 32% on average. Favorite buyers made small but positive returns.

What's the system doing?

1. Building a price model

I use a standard financial formula (Black-Scholes, the same framework banks use to price stock options) to estimate: given where Bitcoin is right now, how likely is it to finish above its starting price by the end of this window?

The key finding: the model is honest. When it says "90% likely," the outcome happens about 90% of the time or more. This is called "calibration."

2. Measuring whether there's actually money to be made

This is where most people fool themselves. The displayed price on the screen is different from the price you'd actually pay if you clicked "buy." The displayed price is like a car's sticker price; what you actually pay (the "ask") is higher.

Near the end of a window, the displayed price might say the favorite is at 77 cents, but if you tried to buy it, you'd pay 93 cents. That 16-cent gap is the spread — it's real, you can see it, but you can't trade through it. My system measures edge against the real price you'd pay, not the fantasy mid-price.

3. Figuring out the right way to trade

There are two ways to trade:

• Taker: you accept someone else's posted price (like buying at the listed store price). You pay a fee (~1.5%) and you need to be fast.
• Maker: you post your own price and wait for someone to accept it (like putting up a yard-sale sign). You pay no fee, get a 20% rebate on fees others pay, and speed doesn't matter.

My finding: taking doesn't work — the fee eats your edge and faster bots get there first. But making on the favorite side might work, because you avoid the fee, earn rebates, and the favorite-longshot bias means buyers are slightly too eager to sell you cheap favorites.

What's confirmed vs. what's still unknown

Why not just do it and see?

Because losing slowly is the worst outcome in trading. If the edge is real but thin (say +2-3%), you might need hundreds of trades to distinguish "genuinely profitable" from "got lucky." Running real money during that learning phase means paying tuition on what might be nothing.

The whole point of this project is the discipline: measure first, commit capital only when the measurement is conclusive.

When does this become real money?

Three things have to happen in order:

1. The favorite-discount has to survive real losses. The system needs to see enough bets where the favorite loses to know the true win rate. If the win rate after losses is still high enough to cover the cost of those losses — gate passed.
2. Maker fills have to be measured. Posting a limit order doesn't guarantee you get filled. And when you do get filled, it might be because something bad is about to happen. This "adverse selection" is being measured with small real orders on a live bot.
3. The combined math has to work. Win rate times fill rate times not-getting-picked-off has to be positive. If it is: start with $200-500 real money and see if reality matches.

What would kill it: the favorites don't win often enough once you count losses (the apparent edge was just small-sample luck), or every time your order gets filled it's because someone smart is dumping on you.

Glossary