TRAD3: turning a pretty reversal chart into code that doesn't lie in hindsight

Not phase 1 → phase 5, but a score

The first naive version was a rigid chain of states: impulse fades → climax → break of level → retest → confirmation. Pretty on paper; in practice half of real reversals never complete that chain — sometimes there's no clear climax, sometimes there's no retest because the trend already reversed without one (a classic V-shaped reversal).

So TRAD3 is built on a mandatory core plus weighted scoring, not a rigid sequence:

• ◆Mandatory: a confirmed break of structure (BOS) and confirmation after it — either a retest with a reversal candle, or an impulsive continuation without a retest.
• ◆Optional, by points: impulse fading before the break (+2), climax on volume/volatility (+2), a break with expanding volume (+1), RSI/OBV divergence (+1), a clear rejection on the retest (+2).

A signal only flies to Telegram once the score clears a threshold — and the message itself shows what that score is made of.

Not “the system said enter,” but “here are five components, three fired — judge for yourself.”

A pure function instead of a stateful machine

The key requirement across the whole system: swing points and levels must never move retroactively. A swing high only counts as confirmed once price has pulled back one and a half to two ATR from it — the system deliberately sees structure with a delay of a few bars. That's the price for a backtest that can't lie: what the system showed yesterday won't change today.

Technically this simplified the architecture more than expected: instead of a murky stateful automaton with persistent state, I ended up with a pure function that recomputes the entire phase picture from scratch on every tick, over a rolling window of ~400 closed bars. Since past swings are never repainted by definition, recomputing from zero gives the same result as an incremental update — without a whole class of bugs like “forgot to reset a counter after a restart.” The only thing that needs to persist between ticks is the fact that a given signal was already sent — a plain UNIQUE constraint in SQLite keyed on symbol, tf, ts_of_break, entry_type.

Two containers, one shared volume

Binance (ccxt, REST) ──► detection.py (свинги, BOS, ретест, скоринг)
                              │
                              ├─► SQLite (WAL): состояния машин, лог сигналов, MAE/MFE
                              ├─► mplfinance: PNG-график с разметкой
                              ├─► Telegram-бот (aiogram): алерт с кнопками 👍/👎/🗑
                              └─► веб-панель (FastAPI): та же история, с графиками

Two Docker containers — the engine and the web panel — share one SQLite volume. The panel deliberately isn't on Next.js: it's a read-only showcase over the same database, it doesn't need a second Node process, FastAPI + Jinja2 is enough. Sign-in is via the Telegram Login Widget, access is currently limited to my own account.

I also factored out a source-agnostic Candle(symbol, tf, ts, o, h, l, c, volume, source) abstraction. Right now the only source is Binance, but once forex through MT5 happens (that physically needs a Windows terminal), the detection core won't care where a bar came from. The difference between “crypto now” and “forex later” is just the implementation of one Feed class, not a rewrite of the engine.

Binance doesn't only trade crypto

On the first run, tickers like XAU/USDT, CL/USDT, and SKHYNIX/USDT snuck into the top-25 by volume — gold, oil, a South Korean stock. Turns out Binance USDM Futures has long sold not just crypto perpetuals but “TRADIFI” perpetuals on commodities and equities, and by volume they comfortably compete with altcoins. Had to filter by the underlyingType == "COIN" field in market metadata.

A fun side effect: since gold and oil are there as genuine exchange-traded crypto perpetuals, I simply added XAU/USDT and CL/USDT straight to the watchlist — no Windows VPS or MT5 needed, those are only required for real forex pairs like EURUSD.

Three directions

• ◆Offline calibration of thresholds against accumulated signal history.
• ◆Forex via MT5 — EURUSD/GBPUSD/USDJPY (gold and oil are already covered through Binance).
• ◆Maybe access for a couple more people — Telegram Login already verifies identity, it just needs a wider allow-list.

For now — watching the alert feed and learning whether to trust my own score.