Binance Prediction Up/Down

A scaffold for trading recurring BTC Up/Down rounds: a round scheduler, screening gates, a decision contract that compares your probability against the quote, order and settlement handling, redemption, and two offline evaluation harnesses.

Read this first

Every risk control, stake size and threshold here is a suggestion, not a safeguard.

• This skill ships no strategy. The bundled hook is an example that makes the pipeline run; the numbers it reads are ones you measure and supply. The reference signal documented under "Strategy hooks" nets roughly +0.65% per round against a ±1.9pp error bar — a worked example of the bar, not a configuration to copy.
• A hit rate is not an edge. If the market already prices the state your signal keys on, a 58% hit rate earns zero. Only step 3 of the walkthrough can rule that out.
• Positions can expire worthless; max loss per round is the stake. There is no stop-loss to place — the position resolves itself.

Dependencies

Candles come from the public Binance market-data endpoint (api/v3/klines) — no key, no session.

baw --version
baw cli-check --required-version "$(grep -m1 requiredCliVersion SKILL.md | tr -dc '0-9.')" --json

Not published. This skill lives only where you installed it. binance-agentic-wallet comes from https://github.com/binance/binance-skills-hub; this one does not ship from there yet.

Getting started

Steps 1 and 3 cost nothing and either can kill the idea. Do both before live mode.

1 · Backtest offline — no login, no config, no money.

python3 backtest.py fetch --symbol BTCUSDT --interval 1h
python3 backtest.py eval  --interval 1h --hurdle 0.521

Use the interval of the rounds you will trade (5m/15m/1h/1d) — the same signal measures very differently across them. eval prints a hit_table to paste into the config.

2 · Write the config — the first run writes an empty template and exits.

python3 updown.py                        # writes the template
$EDITOR $UPDOWN_HOME/config.json         # mode: dry-run, expected_slippage_bps: 0

3 · Collect quotes, then test whether the signal is already priced in.

python3 updown.py                        # dry-run: decides and records, places nothing
python3 backtest.py pricedin             # reads it back

 streak  dir     n  implied P(up)  realized P(up)     gap   read as
      3    1    64          0.421           0.428  +0.007   PRICED IN — no edge here

A gap near zero where your signal fires means the market already knows — stop there. Needs a few dozen occurrences: at 1h with min_streak 3 that is several days.

4 · Check the plumbing — tail -f $UPDOWN_HOME/updown.log. Expect long stretches of nothing; the HEARTBEAT tally, not the absence of orders, is how you tell working from stuck.

5 · Write your own hook — rounds.jsonl has probability, both asks, overround, breakeven and edge per round.

6 · Go live — minimum stake until the money path proves itself, then set expected_slippage_bps from the ALL-IN lines your fills log.

Capability map

Round lifecycle

Twelve topics are listed at any time — three symbols × four cadences (5 min, 15 min, 1 hour, daily), each with its own book, fee and price feed. symbol and timeframe are required config; one daemon trades exactly one combination.

enumerate → gates → probability hook (cannot see quotes) → compare to quote
  → guards → quote → place → confirm fill → settlement → redeem

Two differences from a spot pipeline:

• No stop-loss and no exit plan. The position resolves itself; the whole decision is at entry.
• Winning does not pay you. Payouts sit in PENDING_CLAIM until redeemed. The daemon sweeps on a timer; PENDING_CLAIM should be empty in steady state.

Which feed settles your round

Published per topic as variantData.priceFeedProvider — read it, do not assume. It decides which series your backtest must use. The topic also publishes startPrice, so an outcome can be checked independently once the round closes.

Decision contract

The quote is the implied probability. Three branches, and the third is the common one:

Breakeven hit rate is the quoted price plus costs — buying Up at 0.55 with a 2% fee needs better than 57%. Computed and logged every round by the mechanism layer, not left to the strategy.

Cost = topic feeRateBps + expected_slippage_bps. The order's slippage_bps is a tolerance, not a cost; charging it taxes every round for a worst case that rarely happens.

Screening gates

Three hard gates run in the mechanism layer, before the hook is consulted:

Gate 1 does the most work: an observable fact about the book rather than a forecast. Measured live on one BTC 1-hour round: asks 0.55 / 0.46 → overround 1.0%, spread 0.01 — and 3.0% on the same round 22 seconds later. Screen immediately before ordering.

A fourth check runs at sizing rather than screening, because it depends on the stake:

Fair-value deviation and "breakeven ≤ demonstrated calibration" are your hook's job, not the mechanism layer's: both depend on a model the scaffold does not have. They are listed here so the division is explicit, not because the daemon enforces them.

Thresholds ship unset. A number from someone else's venue describes their book.

Two kinds of market

Strategy hooks

estimate_probability(rnd, ctx) -> float | None   # P(round closes UP); None skips
size_position(rnd, ctx, edge)  -> float          # USD; clamped to stake_usd and budget

ctx carries cfg, policy (verbatim), streak, streak_dir, open, now_ms.

ctx contains no quote, deliberately. An estimator that can see the price anchors on it, and the calibration report then measures something that looks accurate while holding no edge. Enforced at the interface, not left to discipline. Out-of-range or non-numeric returns are rejected with an ALERT, not clamped.

What ships, and what is known about the signal it reads

The hook is where you write your strategy. What ships is an example that makes the pipeline run — every judgement and every number in it should be replaced by what your own analysis shows.

The bundled implementation reads one mechanical feature — how long the current run of same-direction candles is — and looks the answer up in policy.hit_table, which you supply. It ships with no numbers, because a table describes one symbol at one cadence and borrowing one describes nothing. backtest.py eval --interval <iv> prints a table ready to paste.

For reference, the same family measured on the settlement series at every listed cadence:

• A broad reversal tendency of ~+3pp is present at 15m, 1h and 1d. That replicates.
• "Longer streak, stronger reversal" does not — it appears in one cadence out of four and shifts with the sample window. One in four is what noise looks like.

Net of measured costs, at streak ≥ 3 on an at-the-money round:

The error bar on the hit rate is ±0.019 — 2.9× the edge. Telling +0.65% from zero needs about 22,700 rounds: 11.5 years at min_streak 3. That is neither "it loses money" nor "it makes money" — it is unverifiable at the rate these rounds occur, and every figure assumes the round is quoted at fair value, which step 3 of the walkthrough exists to check.

Treat the table above as a worked example of the bar, not as a configuration to copy.

What budget_usd does and does not cap

It caps loss, not turnover. A round is allowed only while

realised losses  +  capital currently at risk  +  this stake   ≤   budget_usd

so the worst case — every open round settling to zero — still lands inside the number you acknowledged. Winning stakes return to the pool and the daemon keeps going.

This matters for long runs. An earlier build treated the figure as cumulative turnover and never released it at settlement, which stopped a $50 budget at $1 stakes after 50 rounds whatever the outcome — a halt unrelated to risk, and incompatible with the sample sizes this strategy family needs. If you want to stop after a fixed number of rounds, use max_rounds_per_day.

Mechanism guardrails

Enforced here; hooks cannot bypass them. Values are yours.

Configuration

$UPDOWN_HOME/config.json (default ~/.local/updown). First run writes an empty template.

Nothing that touches capital has a default. The one exception is mode — absent means dry-run, so an omission cannot become live trading.

Required (mechanism)

Required (policy — read by the bundled hook)

Optional — kline_interval (defaults to the round's cadence) · anything else your hook reads

Has defaults (plumbing) — mode · poll_sec · settle_poll_sec · kline_refresh_sec · l1_category · l2_category

Runtime files and evaluation

$UPDOWN_HOME holds config.json · state.json · rounds.jsonl (every decision, including traded ones) · orders.jsonl · outcomes.jsonl · quotes.jsonl · updown.log. rounds.jsonl and outcomes.jsonl join on key.

eval enforces a train/test split, prints confidence intervals rather than point estimates, counts hypotheses tested, and checks dose-response on the full sample. Any non-live run writes quotes.jsonl, recording every book it can read, including ones your gates reject — sampling only tradable books biases the priced-in analysis.

Three columns matter in calibration: your probability, a constant 0.50, and the quote. Failing to beat the quote column means there is no strategy, however good the hit rate looks.

Platform behavior notes

A dead session must never look like a quiet market. Every prediction read needs an authenticated session and sessions expire. Pre-flight probes the prediction endpoint itself (wallet status can answer while the prediction session is dead); at runtime any auth failure is logged ALERT and recorded as that round's skip reason.

order-book takes --marketId, not --marketTopicId. Different identifiers; a topic contains markets.

The minimum order is 1 USDT — quote rejects less.

A placed order can still fail. Observed: place-order returns success, the order settles to FAILED ("Failed to execute the market order") with zero filled. Placement is a receipt. The daemon polls to a terminal status and on failure returns the stake to the budget and drops the round — otherwise a failed order consumes budget forever and leaves a phantom open round.

Settled PnL is in realizedPnl; pnl comes back null. Reading the wrong one books every win as zero and destroys the calibration ledger.

Redemption is account-wide, not daemon-scoped: the sweep claims everything claimable in the account, including positions this daemon never opened. Settlement bookkeeping covers only rounds in its own state.json.

Quotes expire between quote and place; the daemon refetches once.

Resting sell orders can co-reserve the same shares and only release at resolution, after which further sells fail for insufficient position while the market is still open. Keep at most one resting order per position.

Pre-flight checks

Read-only — do not probe by placing an order.

baw wallet status --json
baw wallet chains --json
baw wallet balance --binanceChainId <chain> --json
baw prediction category list --json             # confirm crypto / up-down exists
baw prediction position list --limit 1 --json   # confirm the prediction session is alive

Running

cd <skill-dir>/scripts && python3 updown.py                     # foreground
nohup python3 -u updown.py >> runner.log 2>&1 & disown          # long-running
python3 updown.py --collect-quotes                              # records only

Going live: set mode to live. The run refuses to start unless the ack file contains the current budget_usd — so it cannot be inherited by copying a runtime directory, and raising budget_usd invalidates it. An agent may write it once the operator has confirmed the number.

Exiting does not close open rounds. They settle on their own and still need redeeming.

Keeping it running

This is a foreground process with an internal loop, not a Unix daemon: no fork, no setsid. nohup … & survives a logout, not a reboot. What the process does and does not guarantee:

launchd (macOS) — ~/Library/LaunchAgents/local.updown.plist:

<plist version="1.0"><dict>
  <key>Label</key><string>local.updown</string>
  <key>ProgramArguments</key>
  <array><string>/usr/bin/python3</string><string>/ABSOLUTE/PATH/scripts/updown.py</string></array>
  <key>EnvironmentVariables</key><dict><key>UPDOWN_HOME</key><string>/Users/you/.local/updown</string></dict>
  <key>RunAtLoad</key><true/>
  <key>KeepAlive</key><true/>
  <key>StandardOutPath</key><string>/Users/you/.local/updown/runner.log</string>
  <key>StandardErrorPath</key><string>/Users/you/.local/updown/runner.log</string>
</dict></plist>

launchctl load ~/Library/LaunchAgents/local.updown.plist

systemd (Linux) — ~/.config/systemd/user/updown.service:

[Service]
Environment=UPDOWN_HOME=%h/.local/updown
ExecStart=/usr/bin/python3 /ABSOLUTE/PATH/scripts/updown.py
Restart=always
RestartSec=30
[Install]
WantedBy=default.target

systemctl --user enable --now updown

The supervisor restarts the process; it does not re-authenticate. A session that expires stops all trading until someone signs in again — plan for that separately, and watch for ALERT lines rather than assuming quiet means healthy.

Known limitations

Disclaimer

This is a scaffold, not a trading product.

• The bundled strategy is a measured reference with no demonstrable positive expectancy net of costs, provided to show that the pipeline runs and the harnesses discriminate.
• Every risk control and parameter here is a suggestion, not a safeguard. The daemon runs the checks you configure; it does not judge whether they suffice.
• Nothing here is investment, financial or trading advice.
• Positions can expire worthless. Automated trading can lose money through market conditions you did not anticipate, platform behavior changes, defects in this software, or defects in the strategy you write.
• Provided as-is, without warranty. To the maximum extent permitted by law, no liability is accepted for any loss arising from its use.
• You are solely responsible for what you run, the parameters you choose, compliance with the laws that apply to you, and any funds you commit.

If you are not prepared to lose the full budget_usd, do not set mode to live.