TechnicalEngine

A free, always-on intraday scanner for US equities. Every 5 minutes during market hours it pulls 5-minute bars for the S&P 500 + Nasdaq-100 (506 tickers - the full Nasdaq exchange is ~3,400 names, which free Yahoo data cannot sustain at this cadence), checks five day-trading long setups, and posts BUY alerts to a Discord channel with price, stop, and 2R/3R targets. Every alert is journaled and scored against what the market actually did.

The Discord channel is buy-alerts only (plus a morning "scanner online" heartbeat). Outcomes - including every loss - are tracked in the local dashboard (engine serve) and engine report, not posted to the channel.

This is not financial advice, and it is not a profitable system. See Does it work? - the honest answer, measured over 60 sessions, is no demonstrated edge. The engine is built to measure that truthfully rather than to hide it. Day trading is risky; most day traders lose money.

Does it work?

No. Run python -m engine backtest yourself, or browse it in python -m engine serve. The shipped configuration, over 60 sessions of 5m bars across 506 tickers (2026-04-29 to 2026-07-24) - 5,400 alerts. Expect a median of 99 alerts per day (mean 108, range 48–235; the backtest's headline 90/day is diluted by the first ~10 sessions, which produce nothing while the RVOL baseline fills). The busiest single 5-minute cycle observed was 127 alerts, which posts as 13 Discord messages over ~13 seconds:

Targets are 2R, so breakeven needs a win rate above 33.3%. exp_r is expectancy in R - the only figure comparable across target settings, and the one that has to be positive:

signal	decided	win rate	exp_r	first half	second half
gap_go	107	34.6%	+0.030	34.3%	35.0%
vwap_reclaim	879	32.7%	−0.018	36.3%	28.7%
orb_breakout	1,091	25.3%	−0.151	30.3%	19.4%
ema9_vwap_pullback	1,236	26.3%	−0.186	28.8%	24.0%
momentum_thrust	833	25.0%	−0.190	27.6%	22.6%

Overall: −748R over 5,400 alerts (−0.139R each; 30.8% win in the first half, 23.7% in the second). Three things were tested and each came back negative:

• Wider targets do not rescue it. Sweeping exit multiples over every alert: 1R −226R, 1.5R −200R, 2R −247R, 3R −295R. The win rate falls faster than the payoff rises (2R needs 33.3%, delivers 27.3%), so wider targets lose more. The shipped 2R setting was chosen for the shape of the alert, not for edge.
• No sub-population survives split-half validation. Hour of day, price bucket, and confluence score were all sliced; with n≥100 in both halves, nothing clears the breakeven bar in both.
• Tuning does not transfer. A grid over signal sets × threshold found every configuration in the 50–100 alerts/day band lands at 44.5–45.3% out-of-sample, while first-half figures spread 50.1–51.8%. That spread is noise; selecting on it is exactly how a backtest lies to you, which is why backtest prints a stability table under every other number.

Two signals - hod_breakout (46.7% / 44.5%) and squeeze_breakout (45.1% / 42.0%) - were retired for being worst in both halves, the only non-overfit criterion available. Fixing squeeze_breakout's session-boundary bug made it measurably worse (47.6% → 43.7%): it had been scoring overnight gaps, not volatility coils.

gap_go is the one setup with positive expectancy (+0.030R) that also holds up across both halves (34.3% → 35.0%, the only signal that does not decay). At ~2 alerts/day and 107 decided trades it is far too thin to build on

• interesting, not proven, and a single bad month would erase it. Everything else decays into the second half by 5–11 points, which is one regime effect rather than four independent failures.

Treat the alerts as "go look at this chart now", never as trade instructions.

Signals (long only)

Signal	Structure
orb_breakout	First close above the 30-minute opening-range high
vwap_reclaim	Close crosses back above VWAP after ≥3 bars below
gap_go	≥2% gap up that holds its open and breaks the first bar's high
ema9_vwap_pullback	Uptrend pullback tags EMA9, holds VWAP, bounces
momentum_thrust	3 straight higher closes above VWAP on expanding range

Two more - hod_breakout and squeeze_breakout - are implemented and tested but retired: they measured worst in both halves of the sample. Clear config.INTERVALS["5m"]["disabled"] to put them back under measurement.

Signals describe structure only. Every alert must additionally clear a central gate - relative volume ≥2× its own time-of-day normal, and price above VWAP - and score enough of 6 confirmations (threshold in config.MIN_SCORE, currently 4, chosen for ~100 alerts/day rather than for edge):

• up on the day
• RVOL ≥ 3×
• closed in the top 40% of the bar's range
• more than one signal firing on the same bar
• market up (SPY above its own VWAP at the same bar)
• relative-strength leader (day change beats SPY's)

Stops come from the setup's structure, then are clamped to between 1.0 and 2.5 ATR - raw structural stops ranged from 0.09% (noise) to 5.6% (not a day trade). One alert per ticker per session.

Setup

Requires Python 3.12+ and a Discord webhook (Server Settings → Integrations → Webhooks → New Webhook → Copy URL).

py -3.12 -m venv .venv
.venv\Scripts\pip install -e ".[dev]"
copy .env.example .env         :: paste your webhook URL into .env
.venv\Scripts\python -m engine fetch
.venv\Scripts\python -m engine test-webhook

The fetch step is not optional. Relative volume is measured against the same time of day on the previous 10 sessions, so a fresh clone with an empty store produces zero alerts until it has that much history - the baseline returns NaN rather than a number computed from too few samples. That is by design: a partial window would give a different answer depending on how much history happened to be loaded. The default fetch pulls Yahoo's full ~60 days.

Run

run.bat

Loops during market hours, sleeps otherwise, restarts on crash, and stops on a configuration error. Or run it directly with python -m engine run.

Commands

python -m engine run [--once] [--dry-run]   scan loop; --dry-run never posts
python -m engine fetch [--interval 1h]      refresh a bar store (5m or 1h)
python -m engine backtest [--interval 1h] [--days 200]
python -m engine replay 2026-07-24          print one past session's alerts
python -m engine report [--by score] [--replay]  hit rates from a journal
python -m engine serve [--port 8765]        results dashboard at localhost:8765
python -m engine test-webhook               post dummy alerts to Discord
pytest                                      99 tests, offline, ~60s

Dashboard

python -m engine serve opens a local, read-only dashboard (no accounts, no external requests - the chart library is vendored): cumulative-R equity curve, win-rate tiles, per-signal and stability tables, and a click-through candlestick chart for every alert with VWAP and the entry/stop/target levels drawn on it.

The two backtest tiers

• 5m (backtest) - the real system, full fidelity. yfinance only serves ~60 days of 5m history, but the local store is never pruned, so the sample grows past Yahoo's window every day the scanner runs.
• 1h (backtest --interval 1h --days 200) - the same signal ideas on hourly bars over 200 trading days. This is a proxy for regime coverage, not a measurement of the live 5m system: the opening range becomes the first hour, the pullback EMA shortens to 4 bars, squeeze_breakout is disabled outright (a 7-bar session cannot host a 20-bar coil), and hourly ATR stops are wide enough that most trades end the day with neither target nor stop hit (small "decided" counts). Measured result: win rates hug 50% there too, in every quarter of the sample.

run --once --dry-run is the useful one when the market is closed: it runs the exact live cycle against the last completed bar and prints what it would have posted.

Tuning

• Universe: engine/sp500.txt + engine/nasdaq100.txt, one ticker per line, merged and deduped by config.universe(). Dots become dashes automatically (BRK.B → BRK-B). A few valid index members are commented out because Yahoo serves no intraday bars for them. config.EXTRA_TICKERS (SPY) is fetched for market context and never alerts.
• Thresholds: all of them live in engine/config.py - gates, score components, stop clamps, cadence, retention. Change one, rerun backtest, and read the stability table before believing the result.
• Interval profiles (config.INTERVALS) are the single source of truth for anything that differs between the live 5m system and the 1h backtest tier. The module-level constants are derived from the 5m profile - never hand-set them. A divergence here once silently deleted 19 sessions of irreplaceable history.

Design notes

• $0 stack: yfinance 5-minute bars, parquet storage, plain webhook POSTs. The only secret is your webhook URL, which lives in .env (gitignored). If it leaks, anyone can post to your channel - delete the webhook in Discord and make a new one; there is no key to rotate.
• Yahoo intraday data is delayed, often by 15 minutes, and labels bars by interval start. The engine only ever judges bars whose interval has fully elapsed. This is a setup scanner, not an execution feed.
• Rule-based on purpose. No ML: honest walk-forward evaluation in a sibling project showed intraday ML barely beating a coin flip, and as the backtest above shows, the rules don't beat one either.
• Signal functions and indicators are look-ahead-safe, enforced two ways: each signal is truncated and recomputed bar by bar, and the whole scan is checked to produce identical output whether or not future bars are present. That causality is also what lets the backtest precompute each ticker once - 41 sessions score in ~50 seconds instead of 2.3 hours.
• Outcomes are 5-minute-bar approximations. When 1R and the stop are both touched inside one bar, the order is unknowable at this resolution, so it is reported as its own ambiguous bucket rather than silently booked as a loss. No fills, slippage, or commissions are modeled.