datajure.window — com.github.clojure-finance/datajure 2.7.2

Window function implementations for datajure. Each function takes a column (dtype reader/vector) and returns a column of the same length. These are called per-partition by the expr compiler when processing :win AST nodes in window mode (:by + :set).

each-prior-op-table clj

Binary operators supported by win/each-prior.

scan-op-table clj

Maps op keywords to binary functions for use in win-scan.

win-bfill clj

(win-bfill col)

(win-bfill col opt)

Backward-fill nil values with the next non-nil value (NOCB — the mirror of win-fills). Trailing nils (after the last non-nil) remain nil. Takes the same optional limit (bare number or {:limit n}) with the same partial-fill semantics. NOCB-then-LOCF combos compose by nesting: #dt/e (win/fills (win/bfill :x)).

[nil 1 nil nil 4] -> [1 1 4 4 4]

win-cummax clj

(win-cummax col)

Cumulative maximum. nil values skipped. Leading nils remain nil. [nil nil 5.0 8.0] -> [nil nil 5.0 8.0]

win-cummean clj

(win-cummean col)

Cumulative mean. nil values skipped. [10 20 30] -> [10.0 15.0 20.0]

win-cummin clj

(win-cummin col)

Cumulative minimum. nil values skipped. Leading nils remain nil. [nil nil 5.0 3.0] -> [nil nil 5.0 3.0]

win-cumsum clj

(win-cumsum col)

Cumulative sum. nil values treated as 0. [10 20 30] -> [10 30 60] [nil 20 30] -> [0.0 20.0 50.0]

win-delta clj

(win-delta col)

Difference from previous element: x[i] - x[i-1]. Returns nil for the first element (no predecessor). [10 20 30] -> [nil 10 10]

win-dense-rank clj

(win-dense-rank col)

SQL DENSE_RANK(): 1-based, dense tie method, based on current row order. No gaps after ties. Pre-sorted [30 20 20 10] -> [1 2 2 3]

win-differ clj

(win-differ col)

Boolean: true where value differs from predecessor. First element always returns true (q convention — no predecessor to match). [A A B B A] -> [true false true false true]

win-each-prior clj

(win-each-prior op-kw col)

Apply a binary operator to (f x[i] x[i-1]) for each element. Returns nil for the first element (no predecessor). Nil propagates: if either x[i] or x[i-1] is nil, result is nil.

op-kw must be one of: :+ :- :* :div :max :min :> :< :>= :<= :=

Generalizes win/delta (op=:-) and win/ratio (op=:div), but without the double-casting of win/delta or the zero-guard of win/ratio. Use win/delta or win/ratio directly when those semantics are needed.

Examples: (win-each-prior :- [10.0 20.0 30.0]) -> [nil 10.0 10.0] (win-each-prior :div [10.0 20.0 30.0]) -> [nil 2.0 1.5] (win-each-prior :max [30.0 10.0 50.0]) -> [nil 30.0 50.0]

win-ema clj

(win-ema col period-or-alpha)

Exponential moving average. The smoothing parameter accepts three forms:

• a number >= 1 → treated as period, alpha = 2 / (1 + period)
• a number < 1 → treated directly as smoothing factor alpha
• an options map → {:alpha a} sets alpha directly, or {:period p} sets it via the period formula — self-documenting alternatives to the implicit numeric dispatch. Seeded at first non-nil value. nil values carry forward last EMA. Leading nils remain nil. ema 2 [10 20 30] -> [10.0 16.67 25.56] ema {:alpha 0.18} [10 …] -> same as ema 0.18 …

win-fills clj

(win-fills col)

(win-fills col opt)

Forward-fill nil values with the last non-nil value. Leading nils (before the first non-nil) remain nil. Matches q's fills convention.

An optional limit — a bare number or {:limit n} — carries the value at most n positions into each nil run, leaving the rest of the run nil (partial fill: pandas ffill(limit=n) / mbmisc h.locf; deliberately NOT zoo's all-or-nothing maxgap). Default: unlimited.

[1 nil nil 4 nil] -> [1 1 1 4 4] [1 nil nil 4 nil] {:limit 1} -> [1 1 nil 4 4]

win-grr clj

(win-grr col)

Inverse-hyperbolic-sine growth (mbmisc grr with IHS=TRUE): asinh(x[i]) - asinh(x[i-1]) over the current partition/order. nil for the first element (no predecessor). A run of zeros has zero growth: when both x[i] and x[i-1] are 0, the result is 0.0 (not asinh(0)-asinh(0)). A nil or non-finite operand yields nil.

[1 2 3] -> [nil (- (asinh 2) (asinh 1)) (- (asinh 3) (asinh 2))] [0 0 5] -> [nil 0.0 (- (asinh 5) (asinh 0))]

win-lag clj

(win-lag col offset)

(win-lag col offset opts)

Lag by offset positions. Boundary positions (no history) are nil, or the :fill value from a trailing options map. Source nils that get lagged stay nil. [10 20 30 40], offset=1 -> [nil 10 20 30] [10 20 30 40], offset=1 {:fill 0} -> [0 10 20 30]

win-lead clj

(win-lead col offset)

(win-lead col offset opts)

Lead by offset positions. Boundary positions (no future) are nil, or the :fill value from a trailing options map. Source nils that get led stay nil. [10 20 30 40], offset=1 -> [20 30 40 nil] [10 20 30 40], offset=1 {:fill 0} -> [20 30 40 0]

win-mavg clj

(win-mavg col width)

(win-mavg col width opts)

Moving average over width rows (expanding at start). nil values skipped. Matches q's mavg convention. Optional opts map: {:min-periods n} requires an n-row window before emitting (e.g. width for a non-expanding R-style window). 3 mavg [10 20 30 40 50] -> [10.0 15.0 20.0 30.0 40.0]

win-mdev clj

(win-mdev col width)

(win-mdev col width opt)

Moving standard deviation over width rows (expanding at start). ddof (delta degrees of freedom) sets the divisor n - ddof: ddof=1 (default) — sample sd, matching R's sd and datajure's sd aggregator; ddof=0 — population sd, matching q's mdev. nil values skipped. A window with n <= ddof finite values yields nil (sample sd of a single value is undefined, like R's sd). The optional 3rd arg is either a bare ddof (back-compat) or an opts map {:ddof d :min-periods n} (:min-periods default 1; width for a non-expanding R sd/rollapplyr window). 3 mdev [10 20 30 40 50] -> [nil 7.071 10.0 10.0 10.0] (ddof=1, default) 3 mdev [10 20 30 40 50] 0 -> [0.0 5.0 8.165 8.165 8.165] (ddof=0, q's mdev)

win-mdowndev clj

(win-mdowndev col width)

(win-mdowndev col width opts)

Moving downside deviation over width rows (expanding at start), MAR=0: sqrt(mean(min(r,0)^2)) over the finite values in each trailing window, with the count of finite values as the denominator (PerformanceAnalytics DownsideDeviation, method='full', na.rm). nil/NaN/±Inf are skipped; a window with no finite values (empty / all-missing) yields nil — undefined, no data to deviate from — a deliberate divergence from R's DownsideDeviation (which returns 0), matching datajure's nil-for-undefined philosophy. A window with finite values but no downside returns 0.0. Optional opts map: {:min-periods n} (default 1; width for a non-expanding R rollapplyr window — the roll-dd convention). 3 mdowndev [1.0 2.0 -2.0] -> [0.0 0.0 ~1.1547]

win-mmax clj

(win-mmax col width)

(win-mmax col width opts)

Moving maximum over width rows (expanding at start). nil values skipped. Optional opts map: {:min-periods n} (default 1; width for a non-expanding window). 3 mmax [30 10 50 20 40] -> [30.0 30.0 50.0 50.0 50.0]

win-mmin clj

(win-mmin col width)

(win-mmin col width opts)

Moving minimum over width rows (expanding at start). nil values skipped. Optional opts map: {:min-periods n} (default 1; width for a non-expanding window). 3 mmin [30 10 50 20 40] -> [30.0 10.0 10.0 10.0 20.0]

win-msum clj

(win-msum col width)

(win-msum col width opts)

Moving sum over width rows (expanding at start). nil values skipped. Optional opts map: {:min-periods n} (default 1; width for a non-expanding window). 3 msum [10 20 30 40 50] -> [10.0 30.0 60.0 90.0 120.0]

win-rank clj

(win-rank col)

SQL RANK(): 1-based, min tie method, based on current row order. Ties (equal values) get the same rank, next rank skips. Pre-sorted [30 20 20 10] -> [1 2 2 4]

win-ratio clj

(win-ratio col)

Ratio to previous element: x[i] / x[i-1]. Returns nil for the first element (no predecessor) and nil when the previous element is zero (avoids Infinity propagation in financial data, matching the div0 philosophy). The simple-return idiom (- (win/ratio :price) 1) then yields nil for the observation after a zero-price row, signalling 'exclude this observation' rather than polluting downstream calculations with Infinity.

[10 20 30] -> [nil 2.0 1.5] [100 0 50 100] -> [nil 0.0 nil 2.0]

win-rleid clj

(win-rleid col)

Run-length encoding group ID. Increments when value changes. [A A A B B A A] -> [1 1 1 2 2 3 3]

win-row-number clj

(win-row-number col)

SQL ROW_NUMBER(): 1-based sequential numbering by current row order.

win-scan clj

(win-scan op-kw col)

Generalized cumulative scan: applies a binary op left-to-right across col (like Clojure's reductions). Nil values are skipped (last good value carried). Leading nils remain nil until first non-nil value is found.

op-kw must be one of :+ :* :max :min. Killer use case: cumulative compounding wealth index via :*

Examples: win-scan :+ [1 2 3 4] -> [1 3 6 10] win-scan :* [1.1 1.2 1.3] -> [1.1 1.32 1.716] win-scan :max [30 10 50 20] -> [30 30 50 50]

win-tlag clj

(win-tlag col date-col)

(win-tlag col date-col shift)

(win-tlag col date-col shift opts)

Date-value-aware lag (mbmisc lbd / statar tlag): row i gets the value of col at the row whose date-col equals date[i] minus shift periods — nil when no such row exists. Unlike the positional win/lag, a gap in the panel yields nil instead of silently reaching back to the wrong period. A negative shift is a lead. shift defaults to 1.

date-col values may be numbers (plain subtraction — years, xbar buckets, encoded periods) or java.time temporals, shifted per the :unit option (:day default, :week, :month, :quarter, :year). Exact-match caveat: calendar arithmetic must land on a date present in the data — for monthly panels keyed on month-END dates, tlag on a normalised month column (e.g. an xbar bucket) rather than the raw date.

Dates must be unique within the partition (aggregate duplicates first, e.g. :agg by the keys); duplicates throw a structured :tlag-duplicate-dates error. nil dates yield nil and are never matched against.

#dt/e: (win/tlag :x :year) / (win/tlag :ret :date 1 {:unit :month})