A queued trade should retain permission only while its approval conditions remain unchanged. Before the next session, define which changes automatically expire that permission, including price, spread, position size, available capital, stop distance, correlated exposure, market status, and signal age.
On August 1, 2012, Knight Capital began routing orders after a software deployment had missed one of eight servers. According to the US Securities and Exchange Commission’s administrative order, a reused system flag activated dormant code on that server. For roughly 45 minutes, Knight sent millions of erroneous orders into the US equity market and lost more than $460 million.
The failure involved more than faulty code. An instruction entered a changed environment while the system treated old state as current authority. That same mechanism appears on a smaller scale when a trader approves a setup at night and lets the order remain actionable after the facts supporting it have moved.
Approval belongs to a specific market state
Suppose an AI queues a stock trade at 9:40 p.m. The proposed entry is $50, the stop is $49, and the position size is 100 shares. The planned loss at the stop is therefore $100, before fees and slippage.
By morning, the stock indicates an opening near $53. If the order still buys 100 shares and the stop remains at $49, the distance to the stop has grown from $1 to $4. The planned loss has changed from $100 to $400.
The trader may still want the trade. That requires a new decision. Yesterday’s approval covered a $50 entry with $1 of risk per share. It did not cover every future price at which the order might fill.
This distinction matters for any approval-gated trading system. Human control has little value when permission survives material changes without another review. The approval must attach to stated conditions, not merely to a ticker and direction.
Write the bedtime mandate as expiration rules
A bedtime mandate defines what can change overnight without requiring another approval. Keep the tolerated changes narrow and measurable.
A practical mandate might say:
- Cancel the queued order if the expected entry moves beyond the approved price band.
- Recalculate position size whenever the entry price or stop changes.
- Require review if the spread exceeds the limit used when the trade was approved.
- Expire the signal after a stated time or before the next market session.
- Require review after earnings, regulatory news, an exchange interruption, or another event relevant to the setup.
- Recheck available cash, margin, open orders, and total exposure before execution.
- Reject any order whose maximum planned loss now exceeds the account’s risk limit.
- Never widen a stop automatically to preserve the original position size.
Each rule answers the same question: which facts formed the basis of consent?
“Approved overnight” is too broad. “Approved while entry remains between these prices, the stop remains here, total planned loss stays below this amount, and no defined invalidating event occurs” creates an auditable boundary.
The mandate should also specify what happens after a breach. Cancellation is usually clearer than silent adjustment. A changed proposal can return to the queue with its new entry, size, stop, estimated loss, and reason for recalculation. The trader then approves or rejects the trade that actually exists.
Reapproval should expose the changed numbers
A reapproval prompt needs enough information to support a fresh decision. At minimum, show the previously approved values beside the current ones:
- proposed entry;
- stop price and distance;
- share or coin quantity;
- estimated loss at the stop;
- spread or expected slippage assumption;
- total exposure after execution;
- signal age;
- the rule that triggered reapproval.
This comparison reduces a common error: noticing that price changed while missing that position risk changed faster.
It also supports disciplined rejection. A signal may still satisfy its technical rules while no longer fitting the trader’s account, exposure, or loss limit. When a human should reject an AI trade that meets all signal rules depends on those account-level constraints.
The same logic applies to forgotten orders. An order left open from yesterday can compete with a new position for capital or create exposure the trader no longer intends. David’s overnight bot and forgotten order shows why reviewing active orders belongs beside reviewing new signals.
Make stale authority visible
Knight Capital’s dormant code became dangerous when a current instruction reached a server carrying different state. The scale was extraordinary, but the control lesson is ordinary: permission must fail safely when its supporting state changes.
Before the next overnight session, write one sentence for each variable that may move: “If this changes by more than this amount, cancel or request approval again.” Then test the mandate against three cases: a price gap, a wider spread, and an existing order that reduces available capital.
The final overnight status should be unambiguous: approved under current conditions, waiting for reapproval, expired, or cancelled. No queued signal should wake up carrying permission that belonged to yesterday.
Educational content, not financial advice.
Comments
No comments yet.