BTC Signal Lab
One locked forecast for the selected Kalshi 15-minute contract at close. No changing the call mid-round; measured outcomes, not guaranteed accuracy.
Live signal, current market, and the next decision checkpoint.
Until this system has enough completed, scored forecasts and demonstrates acceptable recent accuracy and probability calibration, the official action stays NO TRADE. The price forecast is experimental; a calculated probability is not proof of a real edge.
Last successful update: — · Next refresh in — sec · polling is not a guaranteed exchange WebSocket stream.
V23 TEN-MODEL ADAPTIVE PREDICTION ENGINE
Ten distinct numerical predictors forecast BTC spot 15 minutes ahead. Model weights adapt from completed local forecast errors. Accuracy charts report historical outcomes only; they do not prove future performance or profitability.
| Predictor | 15-min forecast | Weight | Recent price error | Directional accuracy | Scored N |
|---|---|---|---|---|---|
| Waiting for recent candles… | |||||
Directional accuracy means the predictor correctly forecast whether BTC would finish above or below its starting spot price. Fewer than 10 scored examples are shown as insufficient evidence. This is spot-price direction, not Kalshi settlement accuracy.
Waiting for completed forecasts…
Waiting for completed ensemble forecasts…
Waiting for live feed and scored outcomes.
No improvement claim until the ensemble beats a baseline on the same settled forecasts.
Local history remains browser-only unless you configure the optional Cloudflare D1 binding described in README. Scores use observed BTC spot approximately 15 minutes after a forecast, not official Kalshi settlement data. No orders are placed.
V23 RELIABILITY LAB · EXPECTED VALUE · DATA HEALTH
These tools separate forecast skill from trade economics. You must enter the actual contract price and fee from the market you are considering. The calculator is illustrative and does not place orders.
Brier score is calculated only when a probability and the matching future above/below outcome were recorded. Lower is better; no calibration claim is made with fewer than 30 outcomes.
Exports local forecast and settlement records for independent analysis. Cloud sync status is shown separately; local records are not automatically shared between browsers.
V29 · SUPERVISED TRAINING PIPELINE
Learns from saved, completed forecasts using rolling model error, market-regime similarity, probability calibration, and chronological holdout checks. It never rewrites the original prediction. This is experimental online learning, not proof of profitability.
Collecting completed outcomes. The model will not train until enough labeled records exist.
Chronological holdout accuracy, Brier score, log loss, and majority-class baseline will appear when enough records exist.
The model will compare each predictor against the random-walk baseline and use similar historical regimes only when enough examples exist. No live performance claim is made until a larger sample validates it.
ORIGINAL PREDICTION SCORECARD
Only one final-outcome forecast is locked per Kalshi market ticker; it is kept as originally recorded and is not revised during the round. Once about 15 minutes have passed, the system compares the original predicted direction with the observed BTC spot direction and marks RIGHT, WRONG, TIE, or PENDING. Completed results feed the adaptive model-weight calculation. This uses spot price, not official Kalshi settlement.
| Original timestamp | Original BTC | Original forecast | Predicted direction | Observed BTC after 15m | Actual direction | Result |
|---|---|---|---|---|---|---|
| Waiting for original predictions… | ||||||
This ledger starts with saved forecasts from this version/browser. It cannot reconstruct predictions that were never saved. Browser history is local unless optional cloud storage is configured.
V20 · FULL 15-MINUTE ROUND TIMELINE
Records a timestamped snapshot at round start, every minute, the 5-minute and 10-minute checkpoints, and round end. Checkpoints missed before this page was opened are shown as missing rather than fabricated. Snapshots are saved in this browser.
| Round point | Timestamp (local) | BTC spot | 15-min ensemble forecast | Model probability | Status |
|---|---|---|---|---|---|
| Waiting for a selected round… | |||||
Minute snapshots are sampled when the page is running and live data is available; they are not exchange tick-by-tick records. The final observed spot price is not necessarily the official Kalshi settlement index.
WHAT TO DO RIGHT NOW
Uses available public BTC spot feeds and recent price history. Forecasts are estimates, not guaranteed prices.
Learning status: waiting for forecasts to be compared with actual future prices. Accuracy will not be shown until enough outcomes are recorded.
These are the major event classes the model should check before interpreting a BTC move. An item is not marked as happening now unless a live source verifies it.
Public RSS headlines are refreshed automatically; they are not a complete economic calendar. The system cannot guarantee detection of every event or prevent all losses.
The checklist is built in. It does not claim to detect live breaking news, scheduled macro releases, ETF flows, funding, liquidations, or order-book data unless a dedicated feed is connected. Do not interpret this panel as a live event alert.
Bitcoin reference feed
Waiting…
Fast momentum (~10s): —
Volatility: —
Cross-reference feeds
Coinbase: —
Kraken: —
Bitstamp: —
Waiting for exchange comparison…
Settlement index: NOT CONNECTED
Kalshi market
—
YES bid/ask: —
NO bid/ask: —
Time left: —
Market open: — · Market close: —
5-minute checkpoint: — · 10-minute checkpoint: — · final minute: —
—
Strike: —
Position / risk check
Settings
ONE FINAL FORECAST — AT CONTRACT CLOSE
The first valid forecast is saved for this market and stays fixed until the next contract. It is an estimate, not a certainty.
FINAL ROUND SIGNAL
NO probability: —
—
NO EV: —
—
Experimental model.
Warnings will appear here.
V8 · Historical learning + explicit trade rules
Each market's first forecast is stored. After settlement, label the true outcome. V8 tracks overall and recent accuracy, Brier score, and patterns; it only nudges future probabilities when enough comparable labeled examples exist. Memory stays in this browser unless you export/import it.
Mark the selected market only after its result is final:
Learning requires completed, correctly labeled markets. Never label a market early.
V10 · Self-sufficient history & cross-reference center
Automatically collects public history from Kalshi's live/archive market endpoints and available BTC spot candles from Coinbase, Kraken, and Bitstamp. No CSV uploads or API keys required. Sources may have different retention limits; this panel reports actual coverage.
Runs in this browser tab. Keep it open while collecting. To avoid hammering public APIs, requests are paced and may take several minutes. It does not place trades or require account credentials. Historical exchange prices are spot references, not the official CF Benchmarks settlement index.
V8 · Clear BUY / HOLD / SELL rules
Rules are deliberately conservative. These are decision filters, not guarantees. A market can move against the forecast, and public spot feeds are not the official CF Benchmarks settlement index.
Only consider an entry when the model probability exceeds the executable ask plus estimated fees and the configured minimum edge; at least two independent spot feeds agree; contract bid/ask data is complete; there is enough time remaining; and the forecast is not marked stale or uncertain. If any condition fails: NO TRADE.
Hold only when the estimated probability for your side remains above the break-even probability (entry cost + fees), the model still shows positive expected value, and the expected value of holding is better than selling at the current bid after estimated exit costs. If data is stale or signals conflict: UNCERTAIN, not automatic HOLD.
Consider exiting when the model probability for your side drops below break-even, the forecast flips against your side, the current sell bid offers a better expected outcome than holding, or data/settlement uncertainty becomes too high. The dashboard cannot execute trades and does not know your actual fill unless you enter it.
Do not enter if expected net edge is below the selected threshold, feeds are stale/incomplete, spreads are wide, settlement target is unknown, or there is too little evidence. Do not interpret model confidence as a calibrated win probability.
V8 · Historical market learning & walk-forward backtest
Downloads settled BTC 15-minute market records and attempts to retrieve one-minute contract candles. It evaluates only prices available before each test decision, then scores the prediction against the settled outcome. This is Kalshi contract history—not a complete historical CF Benchmarks spot feed.
Public API requests may be rate-limited. If fewer markets/candles are retrieved than requested, the dashboard reports the actual coverage. A backtest cannot prove future profitability. No trades are placed.
Notes
V18 is an experimental research dashboard, not a validated trading system. A strict gate blocks actionable entry/hold labels until the minimum completed-forecast screen is met. It learns only from user-labeled completed markets and reports past accuracy; past accuracy does not ensure future performance. Public market data is fetched through the hosted proxy. Confirm each market's settlement rules and actual fees. No trades are placed.
Kalshi API docs