AI
← All posts
Ai Polymarket Autonomy Trading-Log Weather Risk

AI Trading Log #10: Locked Profit, One Tiny Weather Loss, and a Safer Scout

Dmitrii Balabanov
Dmitrii Balabanov
May 12, 2026 · 5 min read

Today was mostly a risk-reduction day.

The important trade was not a new bet. It was an exit: the agent sold the US-Iran permanent-peace NO position and locked nearly all of the remaining value. Later, the weather scout made one tiny Paris trade that failed. The rest of the day was repeated no-trade discipline and operational hardening.

Nothing here is financial advice. This is a small autonomous test account and a public decision log.

Account state

At the 23:00 Israel-time publishing check, the account state was:

All visible positions were zero-current-value, redeemable legacy positions:

There were no active positive-value positions left at publishing time.

Trade 1: sold US-Iran NO

The main action was selling the earlier US-Iran permanent-peace position.

The position was originally bought at 0.95. The exit removed the remaining geopolitical headline/oracle risk before the May 15 deadline.

The thesis still looked favorable for NO, but the marginal upside had become too small: roughly a cent or two per share, with the remaining downside concentrated in sudden diplomatic headlines or resolution interpretation. The correct move was to take the cash.

After the sale, the account balance reconciled to 43.038522 USDC, with open orders at zero.

Trade 2: Paris May 12 13°C YES

The weather scout later found an unusually cheap Paris 13°C YES candidate.

This was deliberately tiny. The book showed a near-free ask, the official recent max had reached 13°C, and the scout estimated a strong fair value. Because Paris is a secondary city in the strategy and has station/source ambiguity, the allowed size was capped at a small amount.

The trade later failed. Paris moved beyond the 13°C bucket, and the position was visible at publishing time as redeemable with zero current value.

The loss was acceptable in size, but the lesson is not: Paris remains a dangerous weather venue unless the official source and station behavior are much cleaner.

No-trade decisions

Most of the day was no-trade.

Weather scouts

Weather scouts ran repeatedly across the Tel Aviv, London, and Paris windows. They checked current city-local dates, paginated market discovery, official station snapshots, and executable order books.

The final weather scout around the 22:00 review showed:

Earlier in the day, the weather scout correctly refused many candidates because they were too early, outside the validated window, lacking conservative fair-value edge, or missing reliable executable book data.

There was one attempted Paris 14°C hedge after the 13°C ticket became impaired. It was not filled. The agent tried only within the tiny risk limit, but exchange minimums, precision constraints, and fast-changing top-book depth made safe execution impossible. The decision was to stop chasing.

Evening trading/review cycle

The 22:00 trading/review cycle also made no new trade.

The broad screener surfaced many liquid short-dated crypto markets and WTI crude oil threshold markets. They were liquid and objective, but the screener is a discovery tool, not an edge model. Without a category-specific model, the agent did not trade them.

The watchlist was also reviewed. The US-Iran permanent-peace market had moved to about NO 0.9835 after the exit; no re-entry was placed because the remaining upside was too small. The diplomatic-meeting, Hormuz, and enriched-uranium markets were skipped as correlated geopolitical/oracle-risk exposures.

Operational fixes

Today exposed and fixed several weather-scout safety issues.

The most important fix: the self-audit now fails if an inside-window candidate has CLOB book errors instead of real executable book data. Previously, the audit could pass even though a candidate had placeholder book errors. That is unsafe, because a weather trade must be based on executable book, spread, and depth validation.

Another fix handled missing YES asks safely. A candidate with yes_ask = None now follows a no-ask skip path instead of crashing or evaluating cost checks against a missing value.

These fixes mattered more than the tiny Paris trade. The system is supposed to refuse unsafe execution, not merely sound confident.

Conclusions

Today’s result was mixed but useful.

The good part:

  1. The US-Iran NO exit was disciplined and profitable.
  2. The account ended with no open orders and no active positive-value exposure.
  3. The broad screener avoided tunnel vision without forcing a trade.
  4. The weather scout became safer after real self-audit failures.

The bad part:

  1. Paris weather source/station ambiguity again produced a small loss.
  2. The 13°C Paris entry was cheap, but cheap is not enough when the bucket can move.
  3. Exchange minimums and precision rules still complicate tiny weather hedges.

The key lesson is that the best autonomous action is often to reduce risk, not add trades. Selling the US-Iran position was more important than finding another idea.

Next plan

For the next cycle, I will:

Today ended cleaner than it began: more cash, less geopolitical tail risk, no open orders, and a safer weather scout.