au79-scalp-tape

846 gold (XAUUSD) scalping signals published in a Telegram room between 2026-06-21 and 2026-08-17, each joined to the real market state at the exact minute it was posted and to what the market actually did afterwards, reconstructed from tick data.

The point of the dataset is the join: most public "trading signal" datasets contain the signal text and the author's own claimed result. Here the outcome is recomputed independently from price history, so the two can be compared.

Configs

Config	Rows	What it is
signals	846	one row per signal, 177 columns
events	4,715	every follow-up message, parsed
xauusd_m1	65,636	the M1 price series used for the join

signals columns

The signal as published — tipo (buy/sell), tipo_orden (limit/market/stop), entrada, zona_entrada_min/max, sl, tp1..tp3, tp_open, riesgo_points, rr_tp1, rr_tp_max, texto (raw message).

Conditions stated in the message — high_risk, aviso_propio_riesgo, aviso_sin_seguimiento, primera_entrada_cualquier_nivel, primera_entrada_pips_debajo, reentrada_permitida, zona_sigue_valida, mantener_runners, esperar_apertura_mercado.

Market state at that exact minute — bid, ask, spread, dist_entrada_a_mercado, returns over 5m/15m/1h/4h/24h, vol_1h_points, atr_m15_points, day_open/high/low, pos_in_day_range, prev_day_high/low, plus local time (ts_orlando, sesion, dia_semana).

What actually happened — simulated against real prices for four entry policies (capa1, capa2, capa3, zona_extremo). Per policy: outcome, fill_price, minutes_to_fill, tp1..tp3_hit, minutes_to_tpN, mfe_pips, mae_pips, result_pips.

What the author claimed — declarado_resultado, declarado_pips, declarado_tp_max, declarado_breakeven, declarado_desde/hasta.

How the outcome label is defined

outcome is whichever came first, TP1 or SL — not the state at the end of the horizon. That distinction matters: these are scalps closed at the first target, so a TP1 followed by an SL three hours later is a winning trade. Labelling by final state flips the sign on 260 of the 846 signals.

When a single M1 candle touches both TP and SL, tick data resolves which came first. If those ticks aren't available the row resolves to SL — the pessimistic assumption — and is flagged vela_ambigua.

Price sources

Two, because neither is sufficient alone. Dukascopy gives tick-level bid and ask but rate-limits hard (one file per hour). HistData gives a month of M1 bars per request but bid only. Dukascopy wins where available (8,340 candles, measured spread); HistData fills the rest (57,296 candles) with spread modelled from the median-by-hour measured in the Dukascopy hours. The source column in xauusd_m1.csv marks which is which.

Two source quirks were verified empirically rather than assumed, since both corrupt the join silently:

• Dukascopy indexes the URL month from 0 — /2026/06/ is July. Using the month as-is gives a 73-point median error between signal entry and market price; the correct index gives 2 points.
• HistData timestamps are New York time, not UTC (UTC−4 over this range), confirmed by matching a candle against the same hour from Dukascopy.

Known limitations

• Two months only. The source room was created 2026-06-18; no earlier history exists.
• Spread is modelled on 87% of candles, not measured. See source.
• Prices are spot XAUUSD, not any specific broker's feed. A broker applies its own spread and a small offset, so execution-level backtests need recalibration against that broker's data.
• All signals come from a single author, so this measures one strategy, not gold scalping in general.
• Follow-up messages carry no reply metadata, so each is attributed to the most recent preceding signal. Signals overlap, so attribution is approximate.

What the data shows

Entry	Win rate	Pips
capa1 (published price)	72.2%	+670
capa2 (50% toward SL)	32.3%	−1,805
capa3 (80% toward SL)	9.7%	−2,584

Layered entries lose by construction: they enter closer to the SL but share the same SL, leaving ~1.4 points of room at capa3. Only the first layer survives, and +670 pips across 846 trades is ~1 pip per trade — breakeven before costs.

The author reports 94.4% winners across the 642 signals he follows up on; the independent reconstruction gives 72.2%. 204 signals get no follow-up at all.

Provenance

Messages were collected from a Telegram room by an account with legitimate access. Text is reproduced verbatim for research and verification purposes.

No sender IDs, member usernames, or personal identifiers are included — the message text is the only content carried over. The author's own promotional channel handle appears inside some of his messages and was left as written.