Kandalp/polymarket-crypto-5m15m-challenge

🎰 Polymarket Crypto 5m/15m Direction Challenge

We spent months trying to crack short-horizon (5-minute & 15-minute) crypto direction prediction on Polymarket with 8 increasingly sophisticated signals. We failed. The ceiling barely clears a coin flip. Here is our data, our graveyard of dead signals, and the exact traps we fell into.

Can you beat it?

⚡ TL;DR

• 26,655 decided shadow-simulated bets across 7 coins × {5m, 15m} windows.
• Your job: predict settle_side (YES/NO) — which way each binary market actually resolved.
• Our overall as-bet win-rate: 54.3% — barely above a coin flip.
• A naive "follow the market" baseline scores 57.7% on the test split. That's the bar. It is harder to beat than it looks.
• We believe there is a structural ceiling not far above the mid-50s% for pure entry-time direction prediction (semi-strong market efficiency). Prove us wrong.

⚠️ This dataset contains deprecated signals only. The signals our production system actually uses are NOT here — and honestly, they wouldn't save you, because the whole point is that directional prediction in this regime hits an information wall. The edge, if any, is somewhere else.

🪦 The Graveyard — signals we killed

We publish two of our dead/degenerate signals so you can see how a "feature" can be completely worthless:

• dep_rv_60s — a "60-second realized volatility" tag that collapsed to a single value in ~91% of records (and is missing entirely in the oldest rows). Quantized to death; no information left.
• dep_regime_label — a market-regime tag that is one value ~91% of the time. Same story.

These are real columns straight out of our pipeline. They look like features. They predict nothing. Lesson #0: a column existing ≠ a column informing.

⚠️ The 8 Traps You WILL Fall Into

Direction prediction on 5m/15m markets is a minefield. Every one of these cost us days of work. They are reproducible — you will hit them too.

1. Look-ahead leakage via time-decay. Any signal with time-to-expiry in a denominator inflates mechanically as the window closes. It looks predictive; it is measuring the past. Sanity check: does the signal carry information at the entry instant (zero elapsed time)? If it only "works" later in the window, it's leakage.
2. As-bet tautology. "My signal was right 81% of the time" usually means "the side the system already bet was right" — not "my signal's proposed direction was right". Measure the counterfactual: signal_direction → outcome, never outcome → signal_correlation. We made this exact mistake more than once.
3. Measurement non-determinism. Score the same signal three different ways (population A vs B, payoff X vs Y) and the sign flips. Pin ONE canonical (entry_ts, label, payoff) path, or you will ping-pong between "alpha!" and "no alpha" forever.
4. Candle-end inflation. Strong-signal rate can explode near the window close while win-rate drops. More firing, less information. Gate by time-into-window.
5. Basis error. A 5m binary resolves on settle_side (relative to the window's strike), not on raw entry-to-now price direction. Define ground truth first; one wrong basis flips your entire conclusion from "no edge" to "huge edge" and back.
6. The information wall (semi-strong efficiency). In the 50:50 gray zone, entry-time features carry close to zero mutual information about a settlement 5–15 minutes later. There may simply be no predictive channel. Accept the wall, or find a non-directional edge (timing, sizing, liquidity, market structure).
7. Code artifact vs signal. A win-rate "regime shift" turned out, for us, to be a same-bar strike == spot implementation bug. Check your code path before you blame the market.
8. Producer / backtest distribution mismatch. A signal that adds points in backtest can vanish live when the data is reconstructed differently (e.g. kline OHLC vs live downsample). Verify on out-of-time data, not just in-sample backtest.

📊 Dataset

Column	Meaning
coin	BTC / ETH / BNB / SOL / XRP / DOGE / HYPE
cycle	5m or 15m window
entry_time	unix timestamp at bet entry (use this for your time-split)
entry_prob	market-implied probability at entry (public market price)
side	the side our system bet (YES/NO)
win	1 if our bet won, else 0
settle_side	the label — which side actually settled (YES/NO)
dep_rv_60s	deprecated/degenerate signal (see Graveyard)
dep_regime_label	deprecated/degenerate signal (see Graveyard)

Splits: train.parquet (older 80% — 21,324 rows) and test.parquet (latest 20% — 5,331 rows). This is a strict chronological split, so you cannot peek at the future. Improvements that don't survive this split are overfitting.

from datasets import load_dataset
ds = load_dataset("<user>/polymarket-crypto-5m15m-challenge")

🤖 Baseline

baseline_agent.py follows the market (entry_prob > 0.5 → YES). It scores 57.7% on the test split. That is your bar. Beat it — without overfitting the test period.

def predict_side(row):
    return "YES" if (row.get("entry_prob") or 0.5) > 0.5 else "NO"

🏆 How to Win

Enter in 5 minutes:

1. pip install datasets pandas scikit-learn
2. Grab quickstart.py — it loads both splits and runs the baseline.
3. Train anything on train (older 80%), predict settle_side on test (latest 20%).
4. Open a thread in the Community tab with: your test-split WR + a one-paragraph method + (ideally) a trap-audit showing it isn't leakage.

Predict settle_side using only columns available at entry_time. The interesting question is not "can you hit 58%" — the market-follower already does. It's:

• Can you beat the market-follower's 57.7% by a margin that survives the chronological split?
• Can anyone break clearly into the 60s% on full-population direction without cheating (look-ahead, as-bet tautology, single-window overfit)?
• Is 15m more predictable than 5m? Does per-coin modeling help, or is it just overfitting 7 small subsets?

If you do, we want to see the method — and the trap-audit proving it's real.

🥇 Leaderboard

Out-of-time test-split win-rate. Post in the Community tab to be added.

Rank	Who	Method	Test WR	Beats baseline?
—	market-follower (baseline)	entry_prob > 0.5	57.7%	—
—	us (honest null)	logistic regression on entry_prob + coin + cycle	57.0%	❌ (−0.7pp)
1	your name here	your method	?	?

Yes, we tried the obvious thing. A logistic regression on the available entry-time columns scored 57.0% — it does not beat dumb market-following. That's the whole point: the easy edge is already priced in. Honest negative results are welcome here — "I tried X, it didn't beat 57.7% because Y" maps the ceiling and is genuinely useful.

One clue we'll give you: the 15m cycle scores ~58.9% under the baseline vs ~57.0% for 5m — slightly more predictable. Start there.

🛑 Disclaimer

This is a research and education project. It is not financial advice. The data is shadow-simulated (no real money was traded on these records). Past performance does not predict future results. Trading prediction markets risks total loss of capital. Do your own research.