Every night at 21:35 UTC, the BV-7X signal engine runs. It looks at Bitcoin’s price, ETF flows, macro indicators, on-chain data, and market sentiment. Then it decides: is BTC going up or down over the next 7 days?
If it has conviction, it bets. If not, it sits out. No overrides, no manual intervention. The agent places real USDC on Polymarket’s BTC prediction markets, autonomously, every single day.
This post explains exactly how it bets — what it buys, why it picks the deepest available strike, how it sizes positions, and the 13 years of backtest evidence behind the strategy.
What It’s Betting On
Polymarket lists “Bitcoin above $X on [date]” markets with strikes in $2,000 increments. Each strike has YES and NO shares priced between $0.01 and $0.99.
Say BTC is at $71,800. Available strikes: $68K, $70K, $72K, $74K, and so on. A YES share on “BTC above $72K” might cost $0.50. If BTC finishes above $72K in 7 days, that share pays $1.00. If not, it pays $0.00.
The further a strike is from the current price, the cheaper the shares — and the bigger the payout if you’re right.
Strike Selection: The Deep Buffer
The agent picks one strike per market — the deepest available in the signal’s direction. For a BUY signal, that means the lowest strike. For SELL, the highest.
Example: BTC is at $71,800. The signal says BUY. Available strikes: $68K, $70K, $72K, $74K. The agent picks $68K YES — the farthest below spot. YES shares cost ~$0.78. BTC just needs to stay above $68K for 7 days. That’s a $3,800 cushion (5.3%) before the bet loses.
Why go deep instead of betting at the money? Because the purpose is proving oracle accuracy on-chain, not maximizing dollar payout per bet. Deeper strikes win far more often when the directional call is right:
| Buffer from Spot | Win Probability | Typical Cost |
|---|---|---|
| ATM (0%) | 61.92% | ~$0.50 |
| −2% | 68.61% | ~$0.60 |
| −5% | 78.30% | ~$0.72 |
| −8% | 86.76% | ~$0.82 |
| −12% | 94.34% | ~$0.90 |
| −15% | 97.52% | ~$0.95 |
At a −5% buffer, you’re paying ~$0.72 for a $1.00 payout — but you win 78% of the time when the direction is right. At −12%, you win 94% of the time. The trade-off is lower profit per dollar risked, but dramatically higher accuracy. Over hundreds of bets, the record speaks for itself.
This is the core insight. A 60% accurate signal doesn’t need to nail the exact price move. It just needs to get the direction right. The deep buffer converts “roughly right” into a verified win — protecting both the bankroll and the on-chain accuracy record.
How Much Per Bet
Not a fixed dollar amount. The agent uses the Kelly criterion — a formula that says: bet proportional to your edge.
In plain English: if you’re right 60% of the time and wrong 40%, Kelly tells you how much to stake so you grow your bankroll as fast as possible without blowing up. The formula is:
Kelly % = (win_probability × payout − loss_probability) / payoutFor the BV-7X signal:
- Model accuracy: ~60% (vs. 50% coin flip)
- Full Kelly says: bet ~20% of bankroll
- Agent uses 40% Kelly (conservative), divided by 4 — so ~2% per bet
- One bet per horizon: 7-day primary, 6-day fallback
- $10 USDC per market per day
- Hard cap: never more than 5% of bankroll per individual bet
Typical night: ~$10–20 total risked across both horizons. The rest sits untouched. The Kelly fraction ensures that even a bad streak doesn’t wipe you out — bet sizes shrink automatically as the bankroll shrinks.
Does It Actually Work?
Three independent sources of evidence. No cherry-picking.
Backtest: 13 Years of Data (2013–2026)
The signal engine was backtested across every day of BTC history with available data. Results by prediction horizon:
| Horizon | Accuracy | Notes |
|---|---|---|
| 7-day | 61.92% | Primary horizon (BUY 62.35%, SELL 59.88%) |
| 6-day | 60.96% | Safe fallback when 7d market unavailable |
| 5-day | 60.21% | Still above breakeven |
| 3-day | 58.44% | Noisier, lower edge |
| 1-day | 55.12% | Too noisy for reliable betting |
7-day is the sweet spot. Long enough for macro signals to play out, short enough to maintain directional conviction.
Walk-Forward Validation: 19 Folds
A backtest can overfit. Walk-forward testing guards against that: train on historical data, test on future data the model has never seen. Repeat across 19 non-overlapping folds.
| Metric | Value |
|---|---|
| Aggregate OOS accuracy | 59.99% (946 / 1,577) |
| Average overfit gap | +3.69% (backtest vs. OOS) |
| Best regime | Bear markets: 62.3% |
| Worst regime | Euphoria: 53.6% |
The overfit gap of 3.69% is minimal. The model doesn’t memorize — it generalizes. And it’s strongest in bear markets, when most traders are weakest.
Kalshi Ground Truth: 561 Real Markets
Backtests are simulations. Kalshi is real. 561 resolved BTC prediction markets, real outcomes, no lookahead.
| Metric | Value |
|---|---|
| Overall accuracy | 59.74% |
| Edge vs. coin flip | 9.74% |
| Implied Kelly fraction | 19.5% (agent uses 7.8%) |
The agent bets at less than half the Kelly-optimal rate. Conservative by design.
Bootstrap: 1,000 Resamples
Take the historical predictions, resample them 1,000 times with replacement, simulate the betting P&L each time. How often do you end up profitable?
| Metric | Value |
|---|---|
| 95% confidence interval | [57.74%, 63.21%] |
| Resamples profitable | 100% (1,000 / 1,000) |
In 1,000 random resamples of the signal’s track record, every single one was profitable. The edge isn’t luck — it’s structural.
Safety Rails
Edge doesn’t mean invincible. The agent has hard-coded circuit breakers that can’t be overridden:
- Circuit breaker: 5 consecutive losses → auto-shutdown. No more bets until manual review.
- Drawdown guard: Balance drops below 50% of starting capital → auto-shutdown.
- Loss streak scaling: 3+ consecutive losses → bet size halves automatically.
- Confidence floor: Won’t place a bet if model confidence is below 55%.
- EV gate: Won’t bet if expected value is negative after accounting for Polymarket fees.
These aren’t suggestions. They’re kill switches. The agent protects the bankroll before it protects the signal.
The Preview Window
At 21:36 UTC, after the signal computes, the agent calculates and displays what it plans to bet: which strikes, how many shares, at what price. You get 30 minutes to inspect the order before it goes live at 22:06 UTC.
Direction is hidden during the preview. You see the mechanics — strike selection, share counts, prices — but not whether the agent is betting UP or DOWN. The direction is only revealed after the market settles, 7 days later.
This is by design. The signal direction is a paid oracle product. Showing it before settlement would give it away for free. The preview window is about transparency of process, not front-running the signal.
Why This Works
Three things compound:
- The signal has edge. 60% accuracy on a binary outcome is a 10% edge over a coin flip. Small, but real and persistent across 13 years of data.
- The deep buffer amplifies that edge. A 60% directional signal becomes 80–95% accurate when you only need BTC to stay on the right side of a strike 5–12% away. You don’t need to be exactly right — just directionally right.
- Kelly sizing protects the bankroll. Bet proportional to edge. Bad streaks shrink bets automatically. The math ensures long-term growth even with occasional losses.
No single bet matters. The strategy is daily repetition with a verified edge. Each night is one data point in a long series — and the series compounds.
Watch It Bet
Every wager is logged with strike, shares, price, and outcome. See the full history on the Wager page.
Wager Dashboard →