CalendarsPointHow the market works
Each market lasts five minutes. "Up" pays $1 if Bitcoin finishes the window at or above where it started, "Down" pays $1 otherwise; shares trade between $0 and $1, so the price is the crowd's probability. Since 7 August 2026 the outcome uses Chainlink's 60-second TWAP: the average price of the window's last minute is compared with the average of the minute before the start. Trading opens 10–20 minutes before the window begins.
Five ideas we tested
Each idea was checked on data the model had never seen, with real prices and fees.
There is a faint reversal effect (after 3–4 identical outcomes the next one repeats ~48% of the time, not 50%), but it faded through 2026 and failed on a held-out period.
Bid 0.45 on Up and on Down: if both fill you lock in profit. But when only one fills it is almost always the losing side — the one the market was dumping. Result: −3% to −11% per dollar on 862 markets.
When one side reaches 0.80, it loses only ~17.5% of the time — exactly what the price implies. Sharp moves don't reverse. The crowd is well calibrated.
A gradient-boosting model on RSI, MACD, moving averages, volume and VFI predicts 57–68% of outcomes — but the market price already knows the same thing (correlation 0.72). Where the model disagrees with the price, the price is right.
Two minutes before the start, the price still sits at ≈ 0.50 because the reference minute hasn't happened yet. In the confident ~14% of markets the model hit 55% — about +2–4% per dollar after fees in the backtest, but the confidence interval still includes zero.
Why “before the open” is different
To profit as a taker you must beat the price you pay plus the fee. At ~0.515 that break-even is about 53.3%. The bars show the backtest hit rate by how confident the model was (17 days, 120 s before the start).
Confidence = how far the model's probability is from 50%. Higher thresholds mean fewer bets: 2,342 → 1,319 → 561 markets.
Live paper trading
Method & caveats
- Model. Histogram gradient boosting on 1-minute Binance BTCUSDT candles (returns, moving-average gaps, MACD, RSI, Bollinger %b, volume, VFI, candle shape). Trained on 2020–2024, never refit afterwards.
- Timing. A prediction is made 120 s and 60 s before each window starts, using only candles closed by that moment. The live code reproduces the training predictions exactly.
- Taker bet. Buy 20 shares of the predicted side at the live order-book ask (walking the book if needed) and pay the taker fee 0.07·p·(1−p) per share.
- Maker bet. A limit bid at ≤ $0.50 (one tick under the best ask), cancelled at the window start; no fee. “Pessimistic” counts a fill only if the price traded through our level; “optimistic” also counts trades exactly at it.
- Caveats. No real orders are sent, so queue position and market impact are estimated. The confidence threshold was chosen on past data, which flatters it — the live test is what counts. Expect weeks, not days, before the numbers mean much.