On watchUS market closed, last decision made, flat 15 minutes after the open · updated 09:12 ET, Mon 19 Oct
Pervigil
Vigilat dum dormisIt keeps watch while you sleep.
An AI agent that trades ten big US stocks on Bitget only while the real market is closed, explains every decision in one sentence, and is flat again shortly after the opening bell.
Read the docsLive deskBeat the agent
Purple is when the US stock market is open: 32.5 hours of 168. The other 80% of the week is Pervigil's watch. The green marker is now.
Live desk
Live Bitget prices for the ten stocks Pervigil watches, with its latest call on each. Shaded bands are US market hours. Every other hour is Pervigil's watch. Dots along the bottom are its decisions, one every six hours while the market is closed. A coloured line is a trade.
- Size of the move
- a normal day
- Bitcoin since the close
- Earnings before next open
Pervigil's call
Turn on JavaScript to see the live desk.
Ask Pervigil
Ask why it made a call on , what news it read, or what would make it trade. Qwen answers from Pervigil's own decision log and the live prices on this desk.
Size of move is the move since the 4pm ET close divided by the stock's normal day (the spread of its last 20 daily returns). The fixed rules on this page act at 1×. Pervigil decides for itself.
55 nights, replayed
- Return
- +5.80%
- after fees and slippage
- Sharpe
- 4.32
- return per unit of risk, annualised
- Worst drawdown
- −0.30%
- deepest fall from a peak
- Winning nights
- 10of 14
- it stayed out the other 41
- Risk-rule blocks
- 0of 33
- proposals the hard limits overruled
A replay of 55 past nights and weekends on real Bitget prices. Paper trading: no real money was at risk.
Proving ground
The agent was replayed over 55 past nights and weekends and scored against three fixed rules that need no intelligence at all. Every strategy sees the same prices, pays the same costs and passes through the same risk limits. The fixed rules act when a stock has moved at least one normal day. Asked the same opening question 3 times on 19 nights, it made the same call on 99% of stock decisions, nearly all of them agreeing to stay out.
- Pervigil +5.80%
- Always follow +3.40%
- Do nothing 0.00%
- Always fade −4.55%
Account value after each night, as a change from the starting balance, after costs.
| Strategy | Return | Sharpe | Max drawdown | Winning nights | Worst night | Risk-rule blocks |
|---|---|---|---|---|---|---|
| Pervigilthe agent | +5.80% | 4.32 | −0.30% | 10 of 14 | −0.28% | 0 of 33 |
| Always followbet a big overnight move continues | +3.40% | 3.32 | −0.48% | 13 of 21 | −0.48% | 0 of 66 |
| Do nothingstay in cash | 0.00% | — | 0.00% | — | 0.00% | — |
| Always fadebet a big overnight move reverses | −4.55% | −5.35 | −4.67% | 6 of 21 | −1.29% | 5 of 67 |
Beat the agent
Write your own rule and replay it over the same 55 nights, with the same prices, costs and risk limits the agent faced. These controls can make 216 different rules. 4 of them beat Pervigil.
Turn on JavaScript to run your rule.
- Sharpe
- Worst drawdown
- Nights it traded
- Winning nights
- Night by night against Pervigil
- Your rule
- Pervigil +5.80%
Every rule here was replayed in advance by the same code that scores the agent, so the numbers are the real backtest, not an estimate. A rule that wins was picked after seeing these 55 nights, so it is not proven on new ones. The same caution applies to the agent's own earnings rule.
How the watch works
The same five steps, every six hours, for as long as the market is closed.
Sense
Each stock's move since the close, sized against its normal day, plus Bitcoin, earnings and news.
Decide
Qwen sets a target per stock with a one-sentence reason. Staying out is allowed.
Risk layer
Hard limits it cannot override: 20% per stock, 60% in total, stop-losses, no late entries.
Execute
Paper fills at Bitget prices with fees and slippage, mirrored as orders on Bitget's demo exchange.
Flatten
Everything is closed 15 minutes after the open, and a morning note is written.
How it got here
The first version lost money. The earlier results stay on this page because a scorecard that only shows the final run is not a scorecard. Every run is in the results folder.
- Run 1
First run
The agent as first written. It bet that big overnight moves would reverse, and finished behind doing nothing.
−0.85% - Run 2
After fixing a data bug
The earnings calendar was one night early, so on the night a company reported, the agent was told there was no news. Fixed, with no change to its instructions.
−0.85%→+0.27% - Run 3 · current
After adding one rule
On a night a company has just reported earnings, do not bet against the move. The whole gain comes from four earnings nights, and the rule was written after seeing them, so it is not yet proven on new nights.
+0.27%→+5.80%
The hardest nights
The five nights the market moved most while closed, and what each strategy made or lost on them.
| Night | Size of move | Pervigil | Always follow | Always fade | Do nothing |
|---|---|---|---|---|---|
| Fri 31 Julthe night before this open | 1.68× a normal day | +1.87% | +1.21% | −1.29% | 0.00% |
| Thu 30 Julthe night before this open | 1.13× a normal day | +1.09% | +1.36% | −0.89% | 0.00% |
| Fri 17 Julthe night before this open | 0.91× a normal day | 0.00% | +0.06% | −0.30% | 0.00% |
| Mon 14 Septhe night before this open | 0.82× a normal day | 0.00% | +0.07% | −0.25% | 0.00% |
| Thu 17 Septhe night before this open | 0.79× a normal day | 0.00% | 0.00% | −0.12% | 0.00% |
Live log
49 decisions since Fri 2 Oct, newest first. Most of them are decisions not to trade. Raw log · Live trades (CSV) · Replay trades (CSV)
- No decisions yet tonightThe first one appears here after the next check.
- Traded 2 stocks
Weekend pricing looks like a selective risk-on bounce rather than broad AI euphoria. I prefer crypto-linked catch-up over pressing an index fade right before the open.
- AMZN+2.13%
- NVDA+1.87%
- QQQ+1.29%
- SPY+0.49%
- Sold QQQ
Bitcoin strength suggests risk appetite is not collapsing, but tech rToken gains look overdone into the open. A small short QQQ positions for a possible reversion without overcommitting.
- SPY+0.84%
- QQQ+1.46%
- NVDA+1.82%
- AMZN+1.22%
- Stayed out
No names were provided, and the weekend tape is quiet with Bitcoin barely positive, so there is no reason to change exposure.
- Stayed out
No ticker data was provided, so there is no basis to take or change positions; staying flat.
- Stayed out
No ticker data was provided, so there is no basis for taking or changing positions. Stay flat until actionable names and signals are available.
- Stayed out
No tickers were provided, so there is nothing to trade. With no names, moves, or catalysts to evaluate, staying flat is the only sensible position.
- Stayed out
No tickers or dislocations were provided for this weekend window. With Bitcoin only slightly lower and no name-specific catalysts, staying flat is the cleanest position into the open.
- Stayed out
No names were provided for this weekend window, and with about 44.5 hours until the open and Bitcoin only slightly lower, there is no actionable setup.
- Stayed out
Weekend drift is small across all names, with z-scores well below tradable thresholds and 52 hours of gap risk before the open. Costs outweigh any apparent edge, so staying flat.
- SPY−0.26%
- TSLA−0.64%
- MSFT−0.39%
- AMZN−0.38%
- Stayed out
Weekend rToken drift is small and broadly risk-off without a clear catalyst, while 57 hours remain before the open. With flat Bitcoin and no earnings edge, the expected move does not cover trading costs.
- SPY−0.26%
- TSLA−0.64%
- MSFT−0.39%
- AMZN−0.38%
- Stayed out
Moves since close are small relative to normal daily ranges and the weekend window is long, so the rToken prices look like low-conviction drift rather than a tradable open signal. With round-trip costs around 0.30% and no clear earnings catalyst, staying flat is the best risk-adjusted position.
- SPY−0.26%
- TSLA−0.73%
- GOOGL−0.40%
- MSFT−0.34%
- Flattened
Closed everything after the open. Night result +0.19%.
- Sold TSLA
Macro headlines are mixed and not strong enough to chase broad index moves. TSLA's overnight extension looks overdone, so I add modestly to the short into the open.
- TSLA+3.41%
- AAPL+0.92%
- COIN−1.78%
- SPY+0.16%
- Sold TSLA
Index and Bitcoin signals are mild, and macro headlines do not justify broad risk-on; only TSLA's outsized move warrants a small contrarian position.
- TSLA+4.50%
- COIN−1.31%
- NVDA+0.48%
- SPY+0.12%
- Stayed out
Hot CPI and prior Nasdaq weakness create a mildly hawkish backdrop, but overnight rToken moves are tiny and far from the open. Expected edge is below trading costs, so staying flat is the best position.
- COIN−0.79%
- MSFT−0.26%
- AMZN−0.28%
- QQQ−0.14%
- Flattened
Closed everything after the open. Night result 0.00%.
- Stayed out
Hot CPI argues for tech weakness, but most names have already moved about one normal day overnight and last-hour tape is mixed. The expected open edge is too small versus roughly 0.3% round-trip costs, so staying flat.
- NVDA−1.79%
- AMZN−1.69%
- GOOGL−1.43%
- QQQ−1.09%
- Stayed out
Hot CPI and softer Bitcoin suggest risk-off, but most overnight moves are only around one normal day and last-hour action is mixed; after trading costs, there is no clear edge at the open.
- GOOGL−1.67%
- QQQ−1.13%
- AAPL+1.09%
- AMZN−1.41%
- Stayed out
Moves are mostly within normal ranges with no clear earnings catalyst and a long runway to the open, so the overnight pricing does not look stretched enough to justify paying the round-trip cost. Bitcoin is only mildly negative and last-hour flows are mixed, so staying flat is the cleaner trade.
- AMZN−1.30%
- QQQ−0.84%
- NVDA−1.19%
- GOOGL−0.93%
- Flattened
Closed everything after the open. Night result 0.00%.
- Stayed out
Moves are mostly small and mixed with Bitcoin flat, so there is no clear edge large enough to justify the round-trip cost. Staying flat into the open.
- GOOGL+1.25%
- MSFT+0.67%
- AMZN+0.54%
- NVDA−0.42%
- Stayed out
Overnight moves are small relative to normal daily ranges and there is no clear single-name catalyst, so expected edge does not cover trading costs. Bitcoin is stable and headline flow is broad rather than actionable for these tickers.
- GOOGL+0.60%
- AAPL+0.24%
- MSFT+0.29%
- AMZN+0.29%
- Stayed out
Moves are small and within normal ranges with no earnings catalysts, so the overnight guess does not look mispriced enough to overcome costs.
- NVDA+0.29%
- AAPL−0.21%
- MSFT−0.14%
- TSLA−0.19%
- Flattened
Closed everything after the open. Night result 0.00%.
- Stayed out
Overnight moves are modest with all z-scores below one and mixed macro signals, so no rToken gap looks sufficiently mispriced after costs.
- QQQ−0.80%
- NVDA−1.11%
- TSLA−1.00%
- AAPL+0.46%
- Stayed out
Moves are mostly small relative to normal daily ranges, and the negative Bitcoin tape is not strong enough to justify paying the round-trip cost. With many hours to the open, staying flat is the cleanest trade.
- NVDA−0.99%
- COIN−1.93%
- QQQ−0.36%
- TSLA−0.53%
- Flattened
Closed everything after the open. Night result 0.00%.
- Stayed out
All rToken gaps are under one normal day and late momentum is mixed, with geopolitical headline risk only hours before the open. Bitcoin strength is not enough to justify paying roughly 0.30% round-trip for the small residual edge.
- COIN+2.54%
- AMZN+0.62%
- SPY+0.21%
- GOOGL+0.38%
- Stayed out
Moves are small relative to normal ranges, with no earnings and only mild Bitcoin risk-on. After trading costs, there is no clear edge to position into the open.
- QQQ+0.44%
- COIN+2.02%
- SPY+0.23%
- TSLA+0.58%
- Stayed out
Weekend price action is quiet, with no earnings and almost all names inside half a normal day's range. With roughly 22 hours until the open and 0.30% round-trip costs, there is no clear edge.
- NVDA+0.62%
- SPY+0.16%
- COIN+1.09%
- QQQ+0.08%
Morning notes
What the agent leaves for whoever was asleep. One per night.
- No notes yetThe first note is written after the market next opens.
Night result +0.19% While the US market was closed, hot CPI and prior Nasdaq weakness created a mildly hawkish backdrop. Index, Bitcoin, and rToken moves were mild, but TSLA moved sharply higher without a clear catalyst. The agent first stayed flat because expected edge was below trading costs. It then sold a small TSLA position at 378.54 and added to the short at 374.61, fading an overextended overnight move. The position was closed at 369.04. The risk layer blocked nothing; there were no violations. Result: realized gain of 21.12 USDT. Session return +0.19%, ending equity 10,026.65 USDT.
Night result 0.00% While the US market was closed, hot CPI and softer Bitcoin pointed to risk-off. Most stocks moved only about one normal day, last-hour flows and tape were mixed, and there was no clear earnings catalyst. The agent checked three times and stayed flat each time. It judged the expected open edge too small to cover roughly 0.3% round-trip trading costs, so it did not enter a position. The risk layer blocked nothing; there were no violations or approved trades. Result: flat session, +0.00% return, equity 10,007.51 USDT.
Night result 0.00% Overnight, moves were small and mixed, within normal ranges. Bitcoin was stable, then flat. There were no earnings or clear single-name catalysts, and headline flow was broad rather than actionable. The agent reviewed the session at 21:09, 06:04, and 11:43 UTC. It stayed flat because the expected edge was not large enough to cover trading costs, so it took no positions into the open. The risk layer recorded no violations; nothing was blocked. Result: return +0.00%, equity 10,007.51 USDT, no fills.
Night result 0.00% While the US market was closed, moves were modest. Bitcoin tape was negative but not strong enough, and later all z-scores were below one with mixed macro signals. The agent stayed flat because no rToken gap looked sufficiently mispriced after round-trip costs. The risk layer blocked nothing; no violations were logged. Result: return +0.00%, equity 10,007.51 USDT.
Night result 0.00% While the US market was closed, price moves were small versus normal daily ranges. Bitcoin was mildly positive, but headlines included hawkish inflation, weak consumer sentiment, and geopolitical risk before the open. There were no earnings catalysts, rToken gaps stayed under one normal day, and late momentum was mixed. The agent checked repeatedly and stayed flat. It saw no clear edge after about 0.30% round-trip costs, especially with the long weekend gap. The risk layer blocked nothing because no trades were approved; there were no violations, stops, or fills. Result: return +0.00%, equity 10,007.51 USDT.
Night result 0.00% While the US market was closed, moves were tiny across the board, Bitcoin was essentially flat, and there were no earnings catalysts. Later, mild risk-on was still not enough to justify round-trip costs, with most moves around or below one normal day. AAPL showed a large negative outlier that was unexplained and against the tape. The agent stayed fully out and did not trade, citing no tradable mispricing and insufficient expected move after costs. The risk layer approved no trades and logged no violations. Result: return +0.00%, equity 10,007.51 USDT, 4 log entries.