☀️ Today's Yes

This page = Yes bets (the model's cheap Yes). The dashboard = No bets.

Today's predicted peak vs what Polymarket charges for Yes.

For each city: the models' predicted peak temperature today, next to the live Polymarket Yes price for the bucket that peak lands in — plus the models' own win chance, so you can see at a glance where the Yes looks cheap and where the market has it right. ← Back to the No-bet dashboard

Bankroll ($) ?

Actions

Auto 5 min

Cities

Simple shows the essentials; Advanced adds every column.

🔔 agree ⚠ differ ● recorded ↷ rest of day ▲ floored + / − edge 🎯 buy peak in, still <90¢, market agrees ⇄ near a bucket edge (coin-flip) ⚠ ~~edge~~ phantom — not fillable / market disagrees ▤ $ order-book depth fillable · ✖ dead empty book

Send a bet to your own local trading server. This button hands the top rule-passing proposal to autotrade.py running on your PC (localhost:8799) — it never talks to Polymarket directly from the browser.

ℹ️ About the Username / Password below: this is the lock on your own local server, which you set yourself with python autotrade.py set-password. It is not your Polymarket login, not your wallet, not any exchange account — it only proves to your own PC that it's you sending the alert. Leave both blank if you haven't set a password (single-user machine).

Live trade

Mode Decided by the 🛡️ Control tab, not here. The server stays in TEST until you arm LIVE there; only then does an alert become a real order.

Server username

Server password

Amount ($)

Max price

Status

The button sends the top proposal that passes the same rules as Auto. In test mode it stops safely; in Go live mode it proceeds to the live execution path.

🤖 How the auto-trading flow works

Step 1

Read the signal

The app compares the forecasted bucket with the live Yes price and the real order-book depth.

Step 2

Apply the rules

Each market must pass the same gates as Auto: edge, win%, price, fillable depth, market-agrees, and daily caps.

Step 3

Build a proposal

The app sizes the surviving candidates and ranks them. This is still a proposal, not an automatic order.

Step 4

Send an alert

The button posts the top rule-passing proposal to the local auto-trade server.

Step 5

Execute or stop safely

In test mode the server stops before any real order; in Go live mode it continues to the live execution path.

The human remains the trigger. The app prepares the signal; the server and your safety settings decide whether that signal becomes a real order.

ⓘ Edge & Net edge assume the forecast is right.

🧮 How it's calculated

today's predicted peak vs the live Polymarket Yes price. The forecast comes from the model that has hit best at each station; the recorded high is from Wunderground/NOAA.

Predicted peak — the day's max from the best-performing model at this station (lowest 14-day error there): its own hourly curve, which the ✓ source link opens so you can verify it, floored at the high already recorded at the settle station (▲ marks a floored value). Falls back to Open-Meteo best_match where there's no skill record yet.
Peak: fc vs rec — compares the two sources on when the peak is. Once the recording has peaked, 🔔 agree = both point to the same hour (a strong "it's in" signal); ⚠ = they disagree; "rec —" until the day peaks.
Peak at — the forecast hour of the max (station-local · Norwegian), re-anchored to the clock each refresh: ↷ switches to the rest-of-day peak once the day-max hour passes, then "day over".
Yes price · Pays — the live "Buy Yes" ask on that bucket (last trade if the book is empty); Pays = 1 ÷ price.
Model win% · Edge · Net edge — win% = share of ensemble members whose (floored) high lands in the bucket; edge = win% − price; net edge subtracts a fee/slippage cost. Buckets the recorded high has already passed are marked dead.
Provider detail — the per-city fold shows each forecaster's number + 14-day accuracy at the station (🏆 = most accurate). Informational — it no longer moves the prediction.
Freshness — auto-refreshes every 5 min (Yes prices + recorded high live each cycle). Early in the local day the spread is wide and value is possible; late in the day the market has mostly converged.

Informational only — you place any bets yourself.

❓ Column guide — what each field means

City	The market's city, with its resolution station (airport) in brackets. The two links are the forecast source (Open-Meteo hourly chart) and the settle source (Wunderground / NOAA station) so you can verify both.
Region	Geographic region. Click the header to group cities by region.
Predicted peak	The day's max from the best-performing model at this station (the one with the lowest 14-day error here — its own hourly curve, which the ✓ source link opens), floored at the high already recorded (▲ = floored). Falls back to Open-Meteo best_match if there's no skill record yet. data HH:MM = when the forecast behind it was fetched.
Peak: fc vs rec	Compares the two sources under the city name — Open-Meteo (forecast) vs Wunderground (recorded) — on when today peaks. 🔔 peak agree = once the recording has peaked, both point to the same hour (a strong "the peak is in" signal — the alert). ⚠ = they disagree on the timing. "fc HH:MM · rec —" = the recording hasn't peaked yet. (The full all-model temperature spread is in the per-city detail below.)
Peak at (local · NO)	The forecast hour of the day's maximum, in station-local and Norwegian time. ● recorded = the measured high already tops the forecast, so the peak is in practice set (measured time shown). ↷ = the day-max hour has passed; showing when the rest of the day peaks.
Turndown	Two lines. Top = measured (the fact): the first recorded reading at the settle station that came in below today's recorded high — temperature, local clock time, drop size (↓) and Norwegian time. A peak is only a peak once a later measurement is lower, so this is a fact from the source confirming the top is behind us — 🔔 once the day-peak is genuinely confirmed (declined and past the forecast peak, so a pre-dawn dip doesn't count). That confirmation is the buy signal. measured: no drop yet = a high is recorded but nothing below it yet; measured: no live feed = no settle-station readings for this city. Bottom = pred (the forecast): the first forecast hour below the peak — a heads-up for when to expect the measured drop ("exp" = not reached yet). A forecast, not a fact.
High so far	The highest temperature recorded so far today at the settle station — the number the market resolves on.
Yes bucket	The Polymarket temperature bucket the predicted peak lands in — the bucket a Yes bet backs.
Yes price	The live "Buy Yes" ask on that bucket (falls back to the last trade when the book is empty).
Pays	Payout multiple if it wins: 1 ÷ price. 25¢ pays 4× (a $100 bet returns $400).
Model win%	The estimated chance the bucket wins — now calibrated: the spread comes from the model's real recent error at this station (not the over-confident raw ensemble), and it tightens as the day peaks (a "nowcast": once the recording has topped out, the outcome is near-certain). A 🎯 buy badge appears when the peak is in, the market agrees on the bucket, and the Yes is still 30–90¢. A ⇄ next to the prediction means it sits right on a bucket edge — a coin-flip with the neighbour, so the win% is genuinely less sure than it looks.
Edge	Model win% − Yes price, in points. Positive = the models think it's underpriced. Assumes the forecast is right. A struck-through ⚠ edge is a phantom — it exists on paper but isn't real value: either there's no fillable order book, or the market has settled on a different bucket (so your reading is likely behind). Phantom edges are demoted, never highlighted. The Yes-price cell also flags a ✖ dead or ~ thin book.
Net edge	Edge after a fee/slippage cost (≈2¢, more when the order book is thin). The honest number: a positive edge that turns negative here isn't really worth it.
Stake	A bankroll-aware suggested bet size (¼-Kelly). Set your bankroll above the table.
Market	agree = the models' bucket is also the market's favourite (higher confidence). differ = they disagree.
Status	Where today's peak is relative to the clock: 🔒 locked (recorded high tops the forecast — bucket decided) · 🟠 waiting for lock (forecast peak hour passed, high hasn't topped it yet) · 🔔 ready (peak within ~90 min) · in ~Xh (further out) · day over. The chips above the table count these.
Their clock · last reading	The station's local wall-clock right now, and how long ago its last observation was recorded (⏱ Xm ago) — so you can see how fresh the recorded high is. Turns red past 2h.
Locked at	When the recorded high first topped the whole forecast curve — the moment the bucket is in practice decided (🔒 time · temp · Norwegian time). "still climbing" = not locked yet.
+1h held?	Did the top hold the hour after the locked peak? ✓ survived = readings ~1h later stayed below it · ✗ beaten = matched or exceeded · — = too early.
Lock held?	Has the locked high stood? ✓ held = 2h+ past the peak with temperatures clearly below · settling = past the peak but still at the high · — = too early.
Track record	This station's settled No-bet history from the daily log: win-rate · average price paid · average P&L per $1. From the dashboard's No bets (not this page's Yes bets) — a general read on how the model-vs-market calls settle here. Needs ≥3 settled days.

Auto — propose, then decide — the engine applies your strict rules to today's bets and produces a sized proposal. This browser page never signs an order — it hands the proposal to your local server (autotrade.py), which is the only thing that can place a real order, and only when you have deliberately switched it from TEST to LIVE. The exact flow is:

Analyze the forecast versus the live market and order-book depth.
Filter every market through the same rule gates: edge, win%, price, fillable depth, market-agrees, and daily caps.
Size the surviving candidates with Kelly-style sizing, capped by your max stake and daily budget.
Review the resulting proposal, then use the live-alert button to send it to the local server.
Decide: in TEST mode the server stops safely and sends nothing; in LIVE mode it runs the real execution path (gated by the caps and the STOP kill switch).

⚙️ Rules & assumptions

Preset

⭱ Import

Today's proposal

Executed & logged trades — what, at what price, and why

Every buy the engine actually logged, newest first — straight from the server's ledger. PAPER rows are simulated fills against the real book; LIVE rows are real orders. Each row shows what (city · bucket), side, size & average price, the cost, and why it was taken: the rule in force, the model win%, the net edge, and whether the market agreed.

Server

Server username

Server password

Paper ledger — would the rules have made money?

🛠️ The executor — where the real order actually happens

Real execution is built, but it lives outside the browser, in a local engine you run on your own machine (autotrade.py + polymarket_clob.py). That is deliberate: money-moving is where the legal (Polymarket terms, gambling/finance law) and security (a funded hot wallet) risk lives, so the signing key never touches this page or a repo. What the engine does when a proposal reaches it:

Re-checks every rule independently (passes_rules), then a second, separate safety gate (preflight_live: caps, freshness, kill switch);
places a FOK limit order only via Polymarket's official py-clob-client — never a market order into the book;
reads the signing key from your local environment (a dedicated trading wallet), never from this app;
enforces hard caps (per-bet, per-day) and a STOP kill switch that blocks all real orders the instant the file exists;
stays in TEST mode by default — it only places a real order once you switch it to LIVE, and logs every step to audit_log.jsonl.

Read EXECUTION.md for the full flow and how to arm it safely. Go live only after the paper ledger and the real forward P&L show a positive, walk-forward edge over enough settled days.

Control — what your local engine is doing right now — a live read-out of your own trading server (autotrade.py on localhost). It shows the current mode, every safety gate, and the most recent activity from the audit log. This tab only reads state — it never places an order. Arming LIVE stays a deliberate, separate step.

Server

Server username

Server password

Auto-refresh (5s)

— not connected

🧰 Setup & connection — zero to a real order

Everything needed to go live, in order. Green = done. The steps mirror python autotrade.py doctor.

🔑 Arm LIVE (key 2 of 2) —

A real order needs two keys: key 1 is the base setup (LIVE mode + a signing key + no STOP file), and key 2 is this deliberate, time-limited arming. Arming expires on its own, so the system is never left hot. Type the exact phrase to confirm.

Confirm phrase

Minutes

🧪 Trial buy — prove the plumbing

Places one tiny real order (hard-capped, max ~$3) to prove that signing, USDC allowance and order execution actually work — this is not a strategy bet. It runs behind the same safety gates + arming, but skips the strategy edge rules (an edge is irrelevant to a plumbing test). The result confirms the exact data used and which rules ran.

Yes token_id

Max price

Amount ($)

Confirm

🚀 Trade — send the top rule-pass proposal

Sends the single best proposal that clears your rules (built on the Today tab) to the server. In TEST it stops safely; armed & LIVE it becomes a real order. Same action as the Today tab — mirrored here so you can run everything from one place.

Amount ($)

Max price

💼 Portfolio — positions & P&L

Net position per market, with cost, a live mark-to-market for open positions, and realized P&L once settled. Settle a resolved market (win/loss) to fill its outcome and P&L, or close a position early with Sell.

📖 How going LIVE works — the whole process (read this)

Going from a forecast to a real order is deliberately a chain of gates. Nothing here happens by accident, and every step is recorded. This is the exact sequence:

Step 1 · Signal

The app builds a proposal

On the Today tab, the engine compares the forecast bucket to the live Yes price and the real order-book depth, sizes it (Kelly, capped), and ranks the candidates. This is a proposal only.

Step 2 · Rules

It must clear every rule

The top proposal must pass the same gates as the Auto tab: net edge, win%, max price, fillable depth, market-agrees, and the daily caps. If it fails, nothing is sent.

Step 3 · Send

You send it to your server

The 🚀 Send rule-pass alert button posts the proposal to your local autotrade.py (with your server login). The browser never signs anything.

Step 4 · Two keys

Both keys must be turned

Key 1: base mode is LIVE (POLYMARKET_LIVE_TEST_MODE=0) + a signing key is loaded + no STOP file. Key 2: you have armed above, and the arm hasn't expired. Miss either → it stays TEST and sends nothing.

Step 5 · Order

The order is placed & logged

Only then does the engine re-check every rule and an independent preflight (caps, freshness, kill switch, arm still active), place one FOK limit order, and write a complete row to the ledger — visible in the Auto tab's trade log.

To go live, in order: ① set a server password (python autotrade.py set-password) → ② set base LIVE in .env (POLYMARKET_LIVE_TEST_MODE=0) and load your PK → ③ restart the server → ④ here, type ARM LIVE and arm for a few minutes → ⑤ send the alert from the Today tab. To stop instantly: click Disarm, or create an empty STOP file in the server folder. Everything — arm, disarm, every alert, every order — is written to audit_log.jsonl and the trade ledger.

Recent activity

Straight from audit_log.jsonl — every alert, login denial, lockout and order, newest first.

ℹ️ What each safety gate means

Can trade live — the bottom line. Real orders are only possible when this is YES: mode is LIVE, a signing key is loaded, and the STOP kill switch is off.
Mode — TEST means nothing real is ever sent; LIVE means orders can go out (still subject to the caps below).
Key loaded — whether a signing key (PK) is present. The key itself is never shown or sent.
Kill switch — a STOP file in the server folder blocks all real orders instantly.
Auth — hash (good), plaintext (works, upgrade with set-password), or off (no lock).
Caps & spent today — the hard per-bet and per-day USD limits, and how much real money has already gone out today.

Covers every location Polymarket runs a daily temperature market for (51 cities, grouped by region — a few, like Jinan & Zhengzhou, have no current daily market and show "no open market"). "Today" = your local date; for Asian cities late in their local day the market is often close to resolved (or already closed). Data: Open-Meteo ensemble + 6 models + yr/MET Norway (+ NWS for US cities), live Polymarket Gamma API, and the recorded high from each market's named resolution source: Wunderground for 44 of the 51 cities, NOAA for Moscow, Istanbul & Tel Aviv (read from the NOAA/IEM feed), and the HK Observatory's own reading for Hong Kong (no live floor there). Yes price = live best ask (falls back to last trade when the book is empty). A Yes bet on one bucket is a narrower, riskier bet than the dashboard's No bets — the models must get the exact bucket right. Informational only, not betting advice.

☀️ Today's Yes

Read the signal

Apply the rules

Build a proposal

Send an alert

Execute or stop safely

Per-city bucket ladders — Yes price vs model odds

Today's proposal

Executed & logged trades — what, at what price, and why

Paper ledger — would the rules have made money?

Close position (SELL)

The app builds a proposal

It must clear every rule

You send it to your server

Both keys must be turned

The order is placed & logged

Recent activity