SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillstv-ta

tv-ta

Frontend & websites

View repository

Updated 10-16 06:19

About this skill

tv-ta — TradingView technical analyst

This skill reproduces the crypto-watch checklist (check.html, 28 indicators) from numbers. It downloads Binance klines (the same BINANCE:{SYM}USDT bars the TradingView widgets chart), computes each indicator with TradingView's built-in formulas and defaults, runs one decision node per checklist item, and combines the answers into a verdict and a trade plan. The scripts do all the judging; your job is to run them, present the report, and explain it.

Requirements

  • Python 3.11 or newer, standard library only (on Windows the launcher may be py).
  • Network access to api.binance.com or data-api.binance.vision. No API key.
  • Optional: tvscreener for scripts/crosscheck.py; a TYPESAFE_API_KEY for Jev nodes.

All paths below are relative to this skill's directory. Run commands from there, or prefix the script path with the skill directory.

Run an analysis

  1. Work out the symbol and timeframe from the request. Defaults: BTC, 4h. Timeframes: 15m, 1h, 4h, 1d. Symbols are Binance USDT pairs (SOL means SOLUSDT).

  2. Run:

    python scripts/run.py <SYMBOL> --tf <TF>

    Useful flags: --no-context (skip the 1h/4h/1d comparison), --live (include the still-forming candle, which matches what the chart shows right now), --json (machine-readable), --log (append the node scores to ~/.tv-ta/logs/ for later calibration), --as-of 2026-09-22T16:00 (replay a past moment using only candles that had closed by then, in UTC; use it to check a call against what actually happened).

  3. Show the markdown report as printed: the Section A and B tables, GATES, FINAL, the multi-timeframe table and the trade plan. Do not recompute, re-score or override any item, and do not change the verdict.

  4. After the report, add a short reading of at most five bullet points:

    • which groups drive the verdict (trend, momentum, volume...) and which disagree;
    • whether the other timeframes agree with the requested one;
    • any gate that is reducing trend weights (choppy market, weak ADX);
    • when checklist tally and weighted composite disagree, say that the weights or gates caused it and name the heaviest rows;
    • for WAIT, restate the two trigger levels.

  5. Keep the closing disclaimer. This is rule-based technical analysis, not investment advice.

If the script exits with an error, report the message and stop. Exit code 3 means market data was unavailable (network, or Binance does not list the pair). Never fill in indicator values or verdicts from memory or general market knowledge. An item marked 未核對 (unchecked) had insufficient data. It is excluded from the denominator and is not a WAIT.

Answer a daily Up or Down question

For prediction-market questions of the form "will {SYM} be up or down" over a day (Polymarket or Binance Wallet "Up or Down on "; "明日漲跌", "今天會漲嗎"), run:

python scripts/updown.py <SYMBOL>

The script picks the next question to settle, so you do not need a date from the user. These markets compare the Binance 1-minute close at 12:00 ET on the settle day with 12:00 ET the day before. The script fetches the target price and runs the checklist on the timeframes in profile.toml [updown]. It returns P(Up) as a volatility-based base chance (from the current distance to the target and the time left) plus a technical tilt. For a question that has already settled, pass --as-of <start time in UTC> and then state the actual result next to the answer.

Show the printed tables, including the question page link and the start and end times in the user's local time exactly as printed. The user opens that link to check the question, so name the question by its Binance title and local end time, not by US Eastern time. Ask the user to compare the printed 目標價 with the one the page labels 需超越的價格. Also show the "TradingView widget item-by-item checklist" matrix: every widget indicator from the crypto-watch main page and o/ page, one row each, with its signal on 1H, 4H and 1D. Summarise which widgets agree across timeframes and which conflict, using the evidence tables. --brief drops the evidence. Then give the answer in one sentence: the lean (Up or Down), P(Up), and the main reason. Name the timeframe driving the tilt, or say that the price distance dominates when little time is left. Say that the tilt scale is not calibrated yet, so the number is a lean and not a fair price. Do not advise a bet size.

Then run python scripts/forward_log.py lean <SYMBOL> and add one line with its two leans, calling the second one experimental and not yet validated: the shipped strategy's lean and the learned model's lean (see the forward test below). Do not blend it into P(Up) and do not prefer it.

Run the daily forward test (scheduled job)

A backtest found that the shipped strategy answers the daily Up/Down questions slightly worse than a coin flip (47–50% per coin over five years, seven coins), and that a frozen alternative, the "learned" model (config/forward_model_v1.json, node weights that may be negative), does slightly better but only within noise. Nobody trusts that until it works on days that have not happened yet. This job records both answers every day, and always-Up as the reference, and scores them once each question settles. Your job is to run the command, relay what it prints, and never interpret the model as a recommendation.

When. Once a day, right after the question starts: 12:00 America/New_York, which is 16:00 UTC in US daylight time (mid-March to early November) and 17:00 UTC otherwise; 00:05 or 01:05 Taiwan time. Any time within 3 hours after the start is fine. A row logged later is marked late and is not scored.

What to run (from this skill's directory; use py instead of python on Windows if needed):

python scripts/forward_log.py run

One run does everything: it settles every finished question, logs today's question for BTC, ETH, SOL, BNB, XRP, DOGE and ADA, and prints a scoreboard. It is safe to run again: a symbol already logged for today is skipped and a settled row is never changed. Records are in ~/.tv-ta/forward/forward.csv ($TV_TA_FORWARD_DIR overrides the folder). Pass symbols to log only some (run BTC ETH); pass --json for machine-readable output.

What to do with the result.

Exit code Meaning You do
0 Done Post the printed summary as it is: the counts line, the scoreboard table and the per-symbol line.
3 Market data was unavailable for some symbol or settlement (the failures are printed) Wait 10 minutes and run the same command once more. If it still fails, report the printed failure lines and stop. The next day's run settles anything left over.
2 Bad input or the model file failed its digest check Report the message and stop. Do not repair, recreate or edit the model file.

Rules you must keep.

  • Copy the numbers from the scoreboard. Do not recompute hit rates, round them differently or compare them yourself; the script already prints the ± range and the paired differences.
  • While the scoreboard says the sample is under 200 questions, say the numbers are progress only and not evidence. Never say the learned model "works", "beats" or "loses to" anything from a small sample, and never tell the user to bet on it.
  • Do not edit forward.csv, do not back-fill a missed day by hand, do not change forward_model_v1.json, nodes.toml or any weight, and do not retrain. A missed day stays missing; the scoreboard shows the coverage.
  • The scoreboard's last line carries the review rule. When the 全部 row reaches 180 logged days, tell the user the review point has been reached and show the table. Deciding whether the learned model replaces the shipped strategy is the user's decision, not yours.
  • To answer "what does the learned model say today", run python scripts/forward_log.py lean <SYMBOL> (writes nothing) and quote both leans and the nodes that drive the learned one. To show the running scoreboard without logging, run python scripts/forward_log.py report.

Tune the decision model

Read references/node_design.md before changing anything. The skill's own config/ holds the shipped defaults:

  • config/nodes.toml has one node per checklist item, with its rule, params (thresholds), weight and group. Setting a weight to 0 keeps the row but removes its vote.
  • config/profile.toml holds the verdict thresholds, gates (Noul nodes that scale whole groups), trade-plan parameters, indicator lengths and the optional Jev settings.
  • config/symbols/<sym>.toml holds per-symbol overrides. BTC and ETH ship with values calibrated from a backtest.

Put a user's own tuning in ~/.tv-ta/config/ (or $TV_TA_CONFIG), not in the skill folder. A plugin update replaces the skill folder, but the user folder survives. Files there are patches layered over the defaults, and each one is optional: profile.toml (any profile values), nodes.toml ([nodes.<id>] tables for all symbols) and symbols/<sym>.toml (per-symbol [nodes.<id>] tables). The report's 設定: line lists every layer that was applied. Edit the skill's own config/ only when you are working in the skill's source repository and the user asks to change the shipped defaults.

To try a change, edit the file and rerun with --cache, which reuses the downloaded candles so only the config differs. Then report the composite and the verdict before and after. Make one change at a time. --no-user-config runs the shipped defaults, for comparison. To check the formulas against TradingView, run python scripts/crosscheck.py <SYMBOL> --tf <TF> (needs tvscreener).

references/checklist.md maps every node to its check.html rule and to the numeric definition it uses. Use it when the user asks why an item got its answer.