Sift

Overview#

The common case is one line with no ceremony. Everything else is optional.

close > ema(21)

That is a complete query. It reads the latest bar of every stock in the universe and keeps the ones closing above their 21-day exponential moving average. There is no timeframe to declare, no latest to repeat, and no wrapper.

Compare the same three intents against Chartink, whose syntax makes you restate the timeframe on every single term:

Sift is not Turing complete: no loops, no user-defined functions, no side effects. Every query is statically analysable, which is what lets the editor underline a mistake before you run anything and lets the engine bound the cost of a scan before it starts.

Query shape#

Five clauses, all optional, conventionally in this order.

on <timeframe>
from <universe>
where <condition>
sort by <expression> [asc | desc]
top <n>

A bare condition is a complete query — where is optional. Clauses may appear in any order. Newlines and indentation carry no meaning, and # or // starts a comment that runs to the end of the line.

on daily
from NSE
where close > ema(21) > ema(50)      # a trend in good order
  and volume > 2x avg(volume, 20)
  and delivery_pct > 55
sort by turnover desc
top 25

Fields#

Written bare, with no parentheses. Each maps to one precomputed column.

Sectors & categories#

Closed sets of string values, tested with `is`, `is not` and `in`.

where sector is "Information Technology"
  and pe < 25 and pe > 0

The values are NSE’s official classification — 22 sectors, 12 macro sectors, 58 industries — and the editor autocompletes them after is, so nobody has to remember that the exact spelling is “Oil Gas & Consumable Fuels”. Matching is case-insensitive; the compiler canonicalises onto the official name.

Derivatives#

Per-stock daily aggregates from the NSE F&O bhavcopy, for the roughly 200 stocks with listed futures & options.

from fno where fut_oi_change_pct > 3 and change > 1

Price and open interest rising together is a long buildup — fresh money agreeing with the move, rather than shorts giving up. Open interest is in contracts, summed across all expiries, because contracts are the unit NSE actually publishes.

from fno where pcr_oi > 0.8 and close > sma(50)

Fundamentals#

Valuation, growth and shareholding, as the market knew them on the scan date.

where marketcap > 20000cr and pe < 30 and pe > 0

Fundamentals are point-in-time. A result filed after the 15:30 close becomes visible from the next session, and a restatement counts only from its own filing date — so a scan on any past date sees exactly what a trader could have known that day, never what the filings later became. That is what keeps the hit-rate replay honest for fundamental screens.

where promoter_change_qoq > 0.5 and close > sma(200)

Dividends

where dividend_yield(1y) > 3% and dividend_years(5y) == 5 and dividend_cagr(5y) > 5
sort by dividend_yield(1y) desc

Dividends are counted by ex-date, in rupees per share on the same split-adjusted basis as close, so a yield survives a split. The dividend functions take a number of years rather than bars — dividend(3y), dividend_yield(1y), dividend_cagr(5y) — and any span from one to ten years is derived at scan time from the ex-date record; the one-year forms read the stored columns. A stock that paid nothing reads 0, not NULL, so dividend_yield(1y) == 0 finds the non-payers. The record begins in 2012; a span reaching before it reads NULL rather than a partial total.

Indicators#

Called with a period, over close.

where rsi(14) < 40 and close > sma(200)

Every indicator accepts any period: the listed ones read a precomputed column and anything else is computed at scan time, so ema(37), adx(7) and bb(50, 2.5) all work. Multi-output indicators take their full parameter list — macd(8, 21, 5), supertrend(14, 2) — and written bare, macd() keeps meaning the stored 12/26/9.

The fixed-window price fields work the same way. Every one of them is the stored period of an indicator that takes any window: high_52w is highest_high(1y), return_1m is return(1mo), and highest_high(3y), return(2y) or pct_from_low(6mo) are derived at scan time from the same bars. A period can be written as a duration — lowest_low(5y) — or as a bar count. The dividend functions count years instead; see Fundamentals.

Multi-output indicators

Indicators producing more than one line take empty parentheses and a sub-field. Omitting the sub-field picks the one marked (default) — bb() is bb().mid.

where macd().line crosses above macd().signal
  and close > bb().upper

Window functions#

Rolling aggregates over any period, computed at scan time rather than read from a column.

where volume > 2x avg(volume, 20)
  and close > max(high, 20 bars)[-1]

The window takes a bar count or a duration — avg(volume, 20), max(high, 20 bars) and min(low, 52w) are all valid. The first argument is any expression, so avg(high - low, 10) gives the mean daily range.

Math functions#

One bar at a time, so they compose with everything.

where abs(close - open) > 2% above avg(abs(close - open), 20)

Unlike a window these read a single bar, so one can sit inside a window — avg(abs(close - open), 20) is the mean candle body over twenty sessions. greatest and least take two or more values; the rest take one.

Counting bars#

How many of the last N — not all of them.

where count(volume > 2x avg(volume, 20), 10) >= 3
  and countstreak(close > sma(20), 30) >= 10

These take a condition where everything else takes a value. has been … for N bars demands every bar; this is how you ask for most of them.

Time travel#

Past bars are negative. There is no future.

where open > high[-1] and close > open

Operators#

The usual comparisons and arithmetic, plus chaining.

Chained comparisons

A chain means what it looks like — each neighbouring pair must hold.

where close > ema(21) > ema(50) > ema(200)

That is exactly equivalent to writing the three comparisons out and joining them with and.

Sugar#

Shorthand for the arithmetic other screeners force you to write by hand.

These are not approximations. Each form is checked against its longhand equivalent in the test suite and must select exactly the same stocks.

Event operators#

Crossings, persistence, runs and extremes — the reason the language exists.

Every one of these compiles to a bounded window expression. In a screener without them you would hand-roll the same thing out of offset arithmetic, and get it subtly wrong.

Crossings

where sma(50) crosses above sma(200)

A crossing is defined on two bars: strictly across now, and not across on the bar before. Add a recency window to catch one that happened a few sessions ago.

where rsi(14) crossed above 30 within 3 bars
  and close > sma(200)

Persistence

Whether something has held for a stretch, rather than being true on one lucky day.

where close has been above ema(21) for 10 bars and adx > 25

Monotonic runs

where volume rising for 3 bars and close > close[-3]

Window extremes

is highest in compares against a window that includes the current bar, so it is true exactly when this bar sets the extreme.

where close is highest in 52w and volume > 1.5x avg(volume, 20)

Candlestick patterns#

Detected at build time and queried as a first-class value.

where pattern is bullish_engulfing
  and close within 3% of sma(50)

Add within n bars for recency, or pattern is not … to exclude one. Definitions use proportional tolerances, so they behave the same on a ₹30 stock and a ₹30,000 one.

Numbers & literals#

Indian and Western magnitudes, both native.

where turnover > 10cr and volume > 5L

Durations convert at 252 sessions a year — 21 a month, 5 a week — so 52w, 12mo and 1y are all 252 bars, and max(high, 52w) is exactly the stored high_52w. No term may reach back more than 1260 bars: 5y is the widest window there is, and 6y is refused. As an argument to a dividend function, a duration is read in whole years instead.

Universes#

Which stocks a scan considers, chosen in the toolbar rather than in the query.

from fno narrows a scan to the roughly 200 stocks with listed futures & options — the set the derivatives fields cover.

from fno where fut_oi_change_pct > 10 and close up 2% over 1 bars

Errors#

Every mistake is reported with the reason and the fix, before the scan runs.

The parser and analyser run in the browser as you type — they are pure and need no database — so the editor underlines a problem immediately, and the identical code validates again on the server.

Recipes#

Complete scans worth stealing. Each one runs.

where close is highest in 52w and rel_volume > 1.5

where close within 3% of high_52w and close > ema(50)

where volume > 2x avg(volume, 20)
  and change > 3
  and close > sma(20)

where volume > 2x avg(volume, 10) and change > 5

where close > donchian_upper[-1]
  and volume > 1.5x avg(volume, 20)

where close > bb().upper and rel_volume > 1.5

where close > keltner_upper and rel_volume > 1.5

where close crosses above pivot_r1 and rel_volume > 1.5

The scan library has 142 of these, each testable against a year of history.

Data caveats#

What the newer data can and cannot honestly answer.

Fundamentals are point-in-time: a scan sees the numbers as the market knew them on the scan date. A result filed after the 15:30 close becomes visible from the next session, and a restatement counts only from its own filing date — which is what keeps the hit-rate replay honest for fundamental screens.

Limits#

What a query is compiled against. Crossing one is an error with a span, never a silent truncation.

The lookback ceiling is the one real queries reach, and it is what makes a scan’s cost knowable before it runs: the analyser computes the deepest bar any term needs, so the engine reads that many and no more. min(low, 3y) is 756 bars and compiles; min(low, 6y) is 1512 and does not. A scan that reaches past a year of history reads proportionally more rows, and only that scan pays for it.

Not supported#

Parts of the language spec this dataset cannot honour, and why.

API & MCP#

Everything on this page works without the browser.

A scan is one POST with the Sift source in the body, and the MCP endpoint gives a coding agent this whole reference as a tool — so Claude Code or Codex can write and run scans against your account unaided. Keys are free on every plan.

Read the API & MCP docs