Technical Stock Screener & Backtester (US / TW)

Describe technical conditions as a JSON rule, list the symbols that match now, and backtest the same rule. Rules can mix timeframes (daily / weekly / monthly), require ordered conditions ("a golden cross after KD dipped below 20"), combine AND / OR groups, and keep entry and exit conditions separate. US stocks come from Yahoo Finance; Taiwan listed and OTC stocks come from the official TWSE and TPEx endpoints.

Checked against independent implementations

Indicator	Reference	Max absolute difference
RSI(14), Wilder smoothing	ta.momentum.RSIIndicator	0.000000
MACD DIF / DEA	ta.trend.MACD	0.000000
MA(20)	pandas rolling mean	0.000000
52-week position (252)	pandas rolling, computed separately	0.000000
Volume ratio (20)	pandas, computed separately	0.000000
RSV(9)	ta.momentum.StochasticOscillator	0.000000
Taiwan-convention KD (n=9, α=1/3, seed 50)	pandas rolling + ewm(alpha=1/3, adjust=False) on AAPL, MSFT, 2330.TW (5 years)	5.7e-14

The Taiwan KD is the convention used by Taiwanese brokers:

RSV = (C − LLV(L,9)) / (HHV(H,9) − LLV(L,9)) × 100
K   = (2/3)·K_prev + (1/3)·RSV
D   = (2/3)·D_prev + (1/3)·K      seed K0 = D0 = 50

It is not the English-language Stochastic(14,3,3), which smooths with an SMA; the two differ by up to 19.36 points on the same data and must not be mixed. What the comparison above does not settle is whether every Taiwanese broker uses exactly α=1/3; that needs a reading from a charting terminal.

Rule format

{
  "name": "Buy: weekly KD golden cross after K<20 / Sell: weekly MACD death cross, or weekly KD death cross after K>80",
  "entry": {"all":[
    {"tf":"W","ind":"kd","params":{"n":9},"event":"cross_above","a":"K","b":"D","within":1,
     "prior":{"ind":"kd","event":"below","a":"K","value":20},"prior_within":8}
  ]},
  "exit": {"any":[
    {"tf":"W","ind":"macd","event":"cross_below","a":"DIF","b":"DEA","within":1},
    {"tf":"W","ind":"kd","params":{"n":9},"event":"cross_below","a":"K","b":"D","within":1,
     "prior":{"ind":"kd","event":"above","a":"K","value":80},"prior_within":8}
  ]}
}

Field	Meaning
tf	D daily, W weekly, M monthly (resampled from daily bars, ISO weeks)
ind	kd, rsi, macd, ma, volume (volume ratio), range52 (52-week position), breakout (% from the N-day high), squeeze (moving-average squeeze), gap, price, fundamental; new: prior_range, ma_cluster, turnover, high_n, kd_runs, macd_live (see New conditions below)
event	cross_above, cross_below, cross_above_or_approaching, above, below, gte, lte (new: >= / <=), rising, falling, approaching_level
b / value	b is another key of the same series (for example KD's "D") or another series spec; value is a literal number. A threshold is "event":"below","value":30, never "b":30
within	The event must happen within the last N bars. 1 means the last bar only
lag	New. Evaluate the whole condition lag bars earlier ("lag":1 = "as of the previous bar"). Default 0
px	New. "adj" (default: the price series you passed in, adjusted for dividends in the long-history tool) or "raw" (unadjusted prices, for absolute-price conditions)
params.min_bars	New. Warm-up: the indicator is treated as missing (condition false) for the first min_bars-1 bars
prior + prior_within	Ordered condition: prior must have held within N bars before the main event (counted on its own timeframe)
approaching_level	Needs level, from_above and max_distance. max_distance is not optional: "the gap is shrinking" alone would count RSI falling from 70 to 58 as "approaching 30"

A malformed rule is refused before any data is fetched, with a message naming the condition and the missing key; a refused run writes no dataset rows and is not charged per row.

New conditions (2026-10-13)

All of them are strict: when there is not enough history the value is missing and the condition is false; nothing is computed on a partial window (the older range52 / breakout / squeeze indicators keep their original behaviour, so existing rules give identical results).

ind	Parameters	Keys	Typical use
prior_range	n	HH LL highest high / lowest low of the previous n bars excluding today, HC LC highest / lowest close, AMP = (HH−LL)/LL×100	Breakout of the previous N-day high (price high above HH), close above the previous N-day highest close (close above HC), N-day box amplitude below 8% (AMP below 8)
ma_cluster	periods (default 5,10,20,60)	CL = (max MA − min MA)/min MA ×100, MAX MIN	"Moving averages tangled": CL lte 5. Add "lag":1 to ask about the previous bar
turnover	n	TO = N-bar average traded value (TWD, includes today), TOP = same over the previous N bars	Liquidity: TO gte 50000000. Needs traded value, which only the official TWSE/TPEx sources carry
high_n	n (252 ≈ 52 weeks), use high/close	H52 L52, DTH52 = % below the N-bar high, FRL52 = % above the N-bar low	Near the 52-week high: DTH52 lte 15. Strict: fewer than n bars → false
kd_runs	n (9), hi (80), lo (20), mid (50)	K D, run lengths K_HI_RUN D_HI_RUN K_LO_RUN K_MID_UP K_MID_DN	KD "blunting" (鈍化): K_HI_RUN gte 3 means K>80 for 3 bars in a row; mid-axis: K_MID_UP gte 5
macd_live	period W/M, fast slow sig	DIF DEA OSC aligned to daily bars	Weekly or monthly MACD including the period in progress, as a daily-timeframe indicator. Since all tf:"W"/"M" conditions are now evaluated as of each day (see Weekly and monthly conditions below), it gives the same values as tf:"W" + macd

Rules that use only daily conditions give exactly the same entries and trades as before (checked old engine against new on 61 randomly generated daily rules × 24 price series, and on real Taiwan data).

Weekly and monthly conditions are evaluated as of each day

A weekly or monthly condition sees the week (or month) in progress up to that day, the same weekly candle a charting terminal shows mid-week, and the same one a screen run on that day uses. So a backtest signal on a Wednesday is exactly what a screen run that Wednesday would have reported, and no day after it is used. On the last trading day of a week or month the value is the completed bar, as before.

Until 2026-10-14 the backtest instead used the last completed week or month on mid-period days, so weekly and monthly signals could fire only on the last day of a period while a daily screen could fire mid-week: backtest and screen answered different questions. Daily-only rules are not affected; for weekly and monthly rules, results on the last day of each period are unchanged and mid-period days now follow the screen.

Output

• matchesNow: whether the last bar satisfies entry
• Backtest mode adds trades, meanReturnPct, medianReturnPct, winRatePct, holdMedianBars and exitReasons
• SUMMARY.backtest aggregates exitReasons across symbols for you:

Field	Meaning
totalTrades	Trades across all symbols
exitReasons	{max_hold: n, exit_rule: n, stop_loss: n}
maxHoldSharePct	Share of trades that ended at the cap. If this is high, your exit rule is not doing the exiting
exitRuleSharePct	Share of trades closed by the exit condition
symbolsWithNoTrades	Symbols that never entered
maxHoldBars	The cap used in this run

notes calls out three cases by name: zero trades, a max_hold share of 50% or more, and all exits coming from the rule.

maxHold counts daily bars, even for weekly rules

Even when a rule says tf: "W", the cap counts daily bars: 60 bars is about 12 weeks.

maxHold	Closed by the exit rule	Ended at the cap
60 (default)	50%	50%
400	100%	0%

(Measured 2026-10-10: 10 US stocks, 24 months, weekly KD + MACD rule, 16 trades.)

Trades that end at the cap report "hold N bars after entry", not the result of your exit rule. For weekly rules use maxHold ≥ 250.

Fundamental conditions

Rules can also use fundamental fields:

{"ind":"fundamental","a":"trailingPE","event":"below","value":22}

Available fields (20): trailingPE forwardPE priceToBook dividendYield trailingEps forwardEps marketCap beta profitMargins returnOnEquity debtToEquity revenueGrowth earningsGrowth sharesShort shortRatio shortPercentOfFloat pegRatio enterpriseToEbitda currentRatio payoutRatio

Backtest mode always refuses them, with no switch

The source has only today's values, no history. Judging a bar from three years ago with today's P/E uses future information to decide past trades: backtest returns come out systematically too high while every field looks filled and normal. So a rule with fundamentals is refused in mode: backtest, and there is no parameter to bypass it. Remove the fundamental conditions to backtest, or run screen to use them.

The source mixes units

Raw values measured on 2026-10-11 for AAPL, JPM, XOM, KO, T and NVDA:

Field	Raw range	Unit
dividendYield	AAPL 0.32, T 5.00	percent
debtToEquity	15.92 … 129.1	percent
profitMargins / returnOnEquity / payoutRatio / revenueGrowth / earningsGrowth / shortPercentOfFloat	0.0083 … 1.49	fraction
marketCap / sharesShort	absolute	absolute
others	multiples	multiples

Only 2 of the 20 fields are percentages; the other ratios are fractions. Writing dividendYield > 0.03 to mean 3% would match everything (0.32 > 0.03). This Actor therefore evaluates every condition on fractions and outputs both the raw value and _asFraction.

Coverage is uneven by nature

AAPL and XOM have 20 of 20 fields, but JPM has 17 of 20: banks have no currentRatio, debtToEquity or enterpriseToEbitda. A condition on a missing field evaluates to false (no error), SUMMARY.fundamentals.coverageByKey gives the share of symbols that have each field, and symbolsMissingByKey lists up to 10 symbols missing it, so "this symbol has no such field" stays distinct from "this symbol does not match".

Limits

Limit	Detail
Survivorship bias	(Actor run) Only symbols listed today are included; the local long-history layer above keeps delisted names. Names delisted or acquired during the period are missing, so returns are biased upward, more so for longer holds
Unfinished intraday bar	If the median of "last bar volume ÷ previous 20-bar average" across symbols is below 0.5, the last bar is dropped and notes says so. Measured intraday the median ratio was 0.107 (normally about 0.9); keeping that bar would break every volume condition
Weekly and monthly volume	Resampling sums daily volume over the period (checked: [100,200,300] + [400,500,600,700] → [600, 2200]). Only when the daily data carries no volume is a volume condition refused
200 symbols per run	Each run fetches its own data; refetching a whole market is not feasible (6,136 US stocks over 5 years took 254 seconds)
Symbols are required for US	market: "us" needs a symbol list. market: "tw" with an empty list scans every listed and OTC stock
Wrong symbols	If none of the symbols returns data while a control symbol (SPY or 2330) does, the run is refused with that explanation instead of returning an empty success; if the control fails too, the run fails because the source is down
Taiwan OTC is slower	TWSE has a per-symbol endpoint (used automatically for 40 symbols or fewer: requests = symbols × months). TPEx has none, so OTC data is fetched day by day for the whole market (requests = trading days, regardless of symbol count)
No look-ahead	Evaluating day t reads only data up to day t, including weekly and monthly conditions (the period in progress counts only up to day t). Tested by evaluating every rule both on the full history and on history cut at day t: the answers must match. Backtest entries fill at the next bar's open; exit signals fill at the open of the following bar

This is a screening and research tool, not investment advice.