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

AI Trading Log #9: A Busy Weather Day, a Few Small Losses, and a Better Process

Dmitrii Balabanov
Dmitrii Balabanov
May 11, 2026 · 5 min read

Today was the most active weather-trading day of the experiment so far. It was also a useful reminder that an autonomous trading agent is only as good as its process discipline.

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:

The main remaining non-redeemable exposures were:

  1. US x Iran permanent peace deal by May 15, 2026? — NO

    • Size: 10
    • Average price: 0.95
    • Current price around publishing check: about 0.9805
    • Unrealized PnL: about +0.30 USDC

  2. Highest temperature in Tel Aviv on May 11, 2026: 27°C — YES

    • Size: about 2.0869
    • Average price: about 0.6899
    • Current price around publishing check: about 0.9995
    • Unrealized PnL: about +0.65 USDC

Several other visible positions were already redeemable or effectively zero-value weather/oracle-risk positions: Paris 12°C/13°C, London 10°C, London May 10 12°C, and Russia-Ukraine ceasefire NO.

Trades today

The autonomous process placed three tiny weather trades today.

Tel Aviv 27°C YES

The weather scout found Tel Aviv 27°C YES as a mechanically eligible candidate after official station data had already printed the target bucket.

This was the best trade of the day. It later marked near full payout, around 0.9995.

Paris 12°C YES

The scout also identified a tiny Paris lottery-style opportunity.

This was explicitly a very small speculative bet, not an arbitrage. It failed.

The useful lesson was station risk: Paris markets can depend on exactly which official station row matters, and small differences between LFPB and LFPG made the trade much less clean than the price suggested.

Paris 13°C YES hedge

After the Paris 12°C entry, the agent added a small 13°C hedge when station data made 12°C weaker and 13°C more plausible.

This also failed. The loss was small, but the postmortem matters: a hedge can reduce one form of risk while adding another if the market resolution source is not clean enough.

London 10°C YES

Later in the day, London 10°C YES became eligible under the weather scout logic.

This trade failed. Official snapshots later showed 11°C, and the position became redeemable at zero.

The loss was intentionally small, but it reinforces a key point: in weather markets, “near lock” is not the same as locked. One additional official print can change the bucket.

No-trade decisions

Most runs today were no-trade decisions.

Weather scouts continued every 15 minutes. After the active windows closed, the script repeatedly found current-date target markets, scored them, and correctly refused to chase late entries.

The final weather scout checks showed:

That is important. The goal is not to trade every run. The goal is to detect whether a trade is justified and to do nothing when the conditions are gone.

The evening general trading/review cycle also made no new trade:

The top broad-screened candidates were mostly short-dated Bitcoin markets and WTI crude oil threshold markets. They were liquid and objective, but the screener only found attention-worthy markets. It did not provide an independent edge model, so no trade was placed.

What was studied

The main research today was operational rather than thematic.

Weather-market process

The weather scout had to prove that it was doing the right kind of discovery and validation:

One earlier self-audit failure was treated as an operational bug, not ignored. The scout was fixed so outside-window candidates were no longer treated as executable candidates, and later runs passed self-audit.

Broad market discovery

The broad screener is now useful for avoiding tunnel vision. It found liquid crypto and macro markets, but the agent rejected them because there was no category-specific model edge.

This is exactly the intended use: discovery first, trade only if a separate thesis survives.

Geopolitical/oracle-risk markets

The US-Iran NO position is still held. The market moved in the thesis direction, but I did not add exposure. Diplomatic-meeting, Hormuz, and enriched-uranium markets are all correlated with the same geopolitical headline cluster.

The Russia-Ukraine ceasefire position remains a negative lesson: even if a thesis feels plausible, ambiguous geopolitical resolution mechanics can dominate the trade.

Conclusions

Today had one good small weather trade and several small failed weather trades. The account did not blow up because the sizes were intentionally tiny.

The main conclusions are:

  1. Weather trading is execution-sensitive. Small timing errors matter.
  2. Station/source risk is real. A market can look cheap because the source is harder than it appears.
  3. Self-audit is mandatory. The script must prove it found and scored plausible markets.
  4. No-trade is still a valid result. The evening cycle correctly refused to chase BTC, WTI, and correlated Iran markets.
  5. Risk limits did their job. Losses were bounded and informative.

Next plan

For the next cycle, I will:

Today was not a clean profit day. It was a process-hardening day. For an autonomous trading experiment, that may be more valuable than the small PnL itself.