The same complexity idea on the bare up/down sequence, ignoring move size (sign-binarized). Β· row #20 Β· custom Β· rolling-200 on close Β· Pinto 2025 (arXiv:2504.01974); sign-binarized length-m word complexity
WATCH
Verdict (worst valid cell)
0.94
Max βΟβ vs panel
0.24
Max RΒ²
1.00
Min h_norm
210
Valid cells
201 / 9 / 0
PASS / WATCH / BAN
10 / 10
Slices done
fast
Cost tier
How this verdict was reached
9 cell(s) landed in the 0.85β0.95 WATCH band and none were banned. The worst cell is βΟβ 0.937 vs bar_petrosian_fd. The whole-candidate verdict is the single worst cell (a feature must be orthogonal in every regime, not on average), so the candidate is WATCH.
What this feature measures
In plain terms: The same complexity idea, but on the bare up/down sequence of the price β throwing away HOW BIG each move is and keeping only its direction. How much organised structure lives in the pattern of ups and downs.
Analogy: Strip the chart down to a string of β²/βΌ and ask how patterned that string is, versus a random sequence of ups and downs.
High vs low: HIGH = the up/down pattern is structured. LOW = the up/down pattern is flat or indistinguishable from random.
Tunable knobs (Axis-2): Word length = m=3 (PR#513) and sign (up/down) binarization β both parameter-free. (Binarization is assumed sign-based per Pinto 2025; flagged for a faithfulness check before promotion.)
How we test it
We slide a 200-bar window across real ClickHouse bars and compute this feature at every bar, for each of 21 (symbol Γ threshold) cells across 6 crypto symbols, repeated over 10 historical 2-month windows ("slices") spanning 2021β2026 β 210 independent tests.
In each test we ask one question only: does this feature carry information the live production panel does not already have? (This is Axis-1 orthogonality β no profit/signal claim is made here.)
Three checks per cell: rank-correlation vs the single closest feature (βΟβ), joint reconstructability from the whole panel (RΒ²), and whether the feature actually varies (h_norm).
Read-only ClickHouse Β· against the live panel + the landed Petrosian/Katz fractal features.
What each number means
Max βΟβ (rank correlation): how strongly the feature tracks the single most-similar existing feature, in its worst cell. 0 = unrelated, 1 = a rename. Lower is better. β€0.85 good Β· 0.85β0.95 watch Β· >0.95 banned (duplicate).
Max RΒ² (joint redundancy): how much of the feature can be rebuilt from all existing features together. 0 = fully new, 1 = fully reconstructable. Lower is better. Same thresholds.
Min h_norm (liveness): how much the feature actually moves (normalised entropy of its own values). Near 0 = basically constant / collapsed = useless. Want β₯0.05.
Worst-valid-cell rollup: the candidate inherits the verdict of its worst of 210 cells β deliberately strict, so a feature only PASSES if it is orthogonal everywhere.
Per-slice (regime) orthogonality
Each row is one 2-month historical window. "Worst βΟβ" is the most-correlated existing feature in that regime's worst cell β lower is better.
Slice
Valid
P / W / B
Worst βΟβ
Top competitor
slice012026-01 β 02
21
21 / 0 / 0
0.81
bar_katz_fd
slice032025-07 β 08
21
21 / 0 / 0
0.73
bar_petrosian_fd
slice052025-01 β 02
21
21 / 0 / 0
0.69
bar_katz_fd
slice072024-07 β 08
21
21 / 0 / 0
0.79
bar_katz_fd
slice092024-01 β 02
21
21 / 0 / 0
0.83
bar_katz_fd
slice122023-04 β 05
21
21 / 0 / 0
0.79
bar_katz_fd
slice142022-10 β 11
21
20 / 1 / 0
0.88
bar_petrosian_fd
slice172022-01 β 02
21
21 / 0 / 0
0.69
bar_katz_fd
slice202021-04 β 05
21
16 / 5 / 0
0.92
bar_petrosian_fd
slice212021-01 β 02
21
18 / 3 / 0
0.94
bar_petrosian_fd
Most-similar existing features (the candidate's nearest neighbours)
Which already-shipped features this candidate most often resembles. If one feature dominates here and βΟβ is high, the candidate may be re-deriving it.
bar_katz_fd β most-similar existing feature in 105 cell(s)
bar_petrosian_fd β most-similar existing feature in 68 cell(s)
trade_intensity β most-similar existing feature in 22 cell(s)
Resource utilization
Hard cap: 2 CPU cores (CPUQuota 200%), 2 GB RAM, spare-only β the loop runs at the lowest priority (CPUWeight 10, nice 19, ionice idle) and only starts when β₯2 cores and β₯2 GB are free, so it never competes with production (ClickHouse / the live sidecar).
This sweep: wall time 8m 12s Β· peak RAM 1915 MB of 2048 MB.
Data volume: 6,333,627 bar-rows processed across 210 cells. 28 of 210 cells were sub-sampled to the most-recent 100,000 bars (largest raw cell β 1,058,962 bars). Sub-sampling keeps giant cells tractable and does not change the correlation estimate.
Safety: the sweep completing means it stayed within 2 GB β had it exceeded, the cgroup would have killed only the loop, never the box, and the cap-worked event would be recorded (never raised).
Worth flagging to the operator
π΄ βΟβ 0.937 crosses 0.85 in β₯1 cell (closest existing feature: bar_petrosian_fd) β that cell is WATCH.
βΉοΈ 28 cell(s) sub-sampled to the 100,000-bar cap (estimate unaffected).
βΉοΈ #20 binarization is assumed sign-of-increment (parameter-free); confirm faithful to Pinto 2025 (arXiv:2504.01974) before promotion.