TT Terminal — Trading Guide

How to actually use the 🏓 table-tennis terminal: what each number is, where the real edges are in these leagues, and the discipline rules. Written in the same spirit as the tennis manual's hard lesson — the model is a tool for finding mispriced lines, not an oracle. Every claim below is checkable against tabletennis/site/metrics.json (also shown live in the terminal header).

1. What you're looking at

The terminal covers the high-frequency factory leagues — Setka Cup (UKR), Liga Pro (RUS), Czech Liga Pro, TT Cup (POL), WTT — the same markets the plan targets: matches every 25–30 minutes, best-of-5 races to 11, all day long.

Two engines feed it:

The character residual is the interesting part: per-player traits learned from how they historically deviate from the analytic expectation — clutch, deuce composure, comeback, front-running, fatigue. When it's active you'll see the 🧠 annotation, e.g. character −3% — Duch A. comeback trait (+0.11).

Header health checks, every session: the model tag + held-out accuracy must be showing; if you see ⚠ live poller stale or ⚠ pre-match file >12h old, the numbers on screen are dead — fix the pipeline (§6) before trading anything.

2. How good is the model, honestly

Held-out walk-forward validation, ~24k test rows (never random-split, features from prior matches only — leakage-safe by construction):

Calibration by bucket (predicted vs actual win rate):

Read the last row twice: the model is overconfident on its biggest favourites — exactly where the tennis engine burned us. Practical rule: haircut anything the terminal shows above ~80%, and never take a heavy favourite at short odds just because True P looks huge.

The live layer is honest too: the residual shipped only because it beat the pure analytic baseline out-of-sample (log-loss 0.5522 → 0.5501 over 159k historical game states). It's a small, real improvement — not magic.

3. Reading the Match Centre

Each row: time (or ● LIVE), players (favourite in white), league, and on live rows the score as games · points (e.g. 1–2 g · 7–9 = down a game 1–2, trailing 7–9 in the current game).

• The big % is True P for player 1 — live it updates every ~8 seconds.
• The sparkline is the recent True P path. Choppy = swingy match; a staircase =

one player steadily taking over.

• ▲/▼ ±pp is live True P minus pre-match. This is the single most useful glance

number: it tells you how much the market's opening price is now stale.

With no match selected, the right pane is the Edge Board:

• 💎 LIVE MOVERS — biggest divergence from pre-match. These are where in-play

prices are most likely to lag reality.

• 🧠 CHARACTER RESIDUALS IN PLAY — matches where a player trait is actively

shifting the number. A comeback trait firing while the scoreboard looks lost is precisely the spot recreational money overreacts to.

• 🎯 STRONGEST PRE-MATCH LEANS — high-confidence scheduled picks with fair

odds, for shopping against your book's opening lines.

4. The trade: where TT edges actually come from

There is no Polymarket for these leagues — you're pricing against a bookmaker, and the journal is paper until your book fills are real. The realistic edge sources, in order of quality:

1. Stale in-play lines. Factory-league matches are fast and books reprice

lazily between games. The moment a game ends, compare the book's new line to live True P. A 30pp mover with the book still near pre-match is the trade.

1. Deuce/clutch spots. At 9–9+ the analytic recursion is exact while books

shade to the favourite by reflex. If the residual also likes the underdog's deuce composure, the dog side is systematically a better price.

1. Pre-match line shopping on high-confidence leans. Only in the 60–75% band —

that's where calibration is clean (see §2). Skip the 80%+ leans entirely.

Mechanics, in the edge panel:

1. Select the match, pick your side (the buttons show True P each way).
2. Type the odds your book is actually showing — never leave it on model-fair;

fair odds by definition have zero edge and exist only as a reference.

1. Read the edge line. Floor: +3% for TT (higher than the tennis terminal's 2%

— these books run 6–8% margins and the model is coarser here). Below the floor, don't. The discipline IS the system.

1. Size at ¼ Kelly, hard-capped at 2% of bankroll per bet. The panel shows both

Kelly numbers; TT variance is brutal — a best-of-5 to 11 is a coin that flips fast, and you'll have 40+ opportunities a day. Volume, not size, is where the EV compounds.

1. ⚡ LOG PAPER BET. Every bet, no exceptions — including the ones you'd rather

forget. An unlogged journal is a lie you tell yourself.

5. The Bet Tracker is the product

The tennis terminal's most important discovery came from its journal: the engine's claimed edge was anti-predictive (bigger shown edge → bigger losses). The TT journal exists so we find that out here before real money does.

• Settle every bet honestly (WIN / LOSS / VOID) the moment the match ends.
• Watch P@bet vs your actual hit rate. If bets logged at ~65% are winning ~55%,

the model is overconfident in your selection zone — widen the haircut, or stop.

• ROI over 50+ settled bets is signal. ROI over 10 is noise; do not resize on it.
• One league at a time when starting out. Setka and Liga Pro have different

personalities (pace, walkover rates, motivation patterns) — learn one before spreading.

Hard risk rules (non-negotiable): max 2% of bankroll per bet · stop for the day at −5 units · no bet without a logged edge ≥ +3% · nothing above True P 80% · never chase a loss into the next 25-minute match — there is always another one, which is exactly why tilt is more dangerous in TT than anywhere else.

6. Keeping the data alive (ops)

The terminal reads local pipeline artifacts via /api/tt. Three things must be running:

python sofa_proxy.py                      # :3001 — TLS-impersonating data proxy
python -m tabletennis.pipeline            # daily: ingest → retrain → predictions.json
python -m tabletennis.live                # 8s in-play poller → live_predictions.json

The header tells you when any of these has died (feed unreachable / stale warnings). Retrain roughly daily — Elo drifts fast in leagues where players log hundreds of matches a month, and every state the live poller records grows the training set for the momentum/serve traits that need point sequences.

*Model outputs are calibrated probabilities, not guarantees. These leagues exist to generate betting volume; the books pricing them are not stupid. Bet only what you can afford to lose.*