Function shinoharaIntensityRatio

Shinohara Intensity Ratio (篠原レシオ) — the pair of ratios Shinohara built to read how hard a market is being pushed, each a ratio of two sums over the same period-bar window, as a percent:

${prefix}Strong = 100 · Σ(high − open)      / Σ(open − low)        ← the A ratio
${prefix}Weak   = 100 · Σ(high − prevClose) / Σ(prevClose − low)   ← the B ratio
Copy

Appends two columns. 100 is the neutral reading in both: above it the window travelled further up than down, below it the reverse. The two measure the same push against different reference points — Strong against each bar's own open, so it reads the session's conviction; Weak against the previous close, so it reads how the session behaved relative to where the last one left it, gaps included.

Neither ratio is bounded, and the B ratio can invert

${prefix}Strong's terms are non-negative on any well-formed bar (low ≤ open ≤ high), so it runs from 0 upwards and is unbounded above.

${prefix}Weak's are not: on a bar that gapped up, low > prevClose, so prevClose − low is negative and the bar subtracts from the denominator. On an instrument that gaps more than it ranges, the denominator can therefore approach zero from either side and the reading swings violently or changes sign. That is the definition rather than a defect — the B ratio is asking "how far did the market travel above the previous close, per unit of travel below it", and on a market that never traded below the previous close the answer is genuinely unbounded — but it is worth stating plainly, because it is exactly what the package's own oracle input does: 33 of its 80 bars gap, and ${prefix}Weak there spans −22,761.08 … 7,620.00 against a ${prefix}Strong of 45.51 … 584.50 (scripts/oracle/generate.py, period 26). Read the B ratio on bars whose ranges are wide relative to their gaps, or read the A ratio.

F-AMBIG — which ratio is "strong", and the source followed

The corpus flags this study as F-AMBIG on the A/B conventions, and it is a naming fork rather than an arithmetic one. What ships:

• The arithmetic is the standard Shinohara A and B ratios as they are published in the Japanese technical-analysis literature — A against the open, B against the previous close, both over 26 bars, both ×100.
• The column names come from the corpus' own study list (strong / weak, docs/notes/financial-indicators-assessment-2026-07.md §6.6), mapped A → ${prefix}Strong and B → ${prefix}Weak.

The alternative convention charts them the other way round (B as the "strong" line, on the argument that a gap-relative reading is the stronger signal), and some vendors simply label them A and B and leave the reading to the reader. That matters because the two lines are genuinely different: measured on the package's oracle input at period 26 they sit 23,312.47 apart at their widest (scripts/oracle/generate.py), so a build that swapped the labels would not be wrong by a rounding. A caller who wants the other convention renames the columns; the study does not offer a flag, because a flag would make one study into two.

The wrong turn worth naming is a different one: reading the B ratio against the bar's own close rather than the previous one. That is the mistake that leaves the shape intact, and it sits 22,862.88 away (measured); the generator asserts the separation.

Warm-up — per column, and they differ by one bar

${prefix}Strong reads one bar at a time, so it first prints at period − 1 (bar 25 at the default). ${prefix}Weak reads the previous close, so bar 0 has no term and it first prints at period (bar 26). Each column warms up when it can rather than both waiting for the slower, which is what macd established.

Both sums come from rollingMeanValues, so a window holding a gap emits nothing rather than averaging around it — the rule for a derived array, and the one that keeps "a 26-bar sum" meaning 26 bars.

Edges

• Invariant under scaling AND shifting every price. Every term of both sums is a difference of two prices from the same bar (or the same adjacent pair), so a shift cancels inside each term and a scale factor cancels between numerator and denominator. That is a stronger statement than most ratio studies get, and both halves are pinned as property tests — a build that dropped the 100 · factor would still pass both, so the unit tests pin the scale.
• A zero denominator → undefined, not 0 or Infinity. Σ(open − low) = 0 on real bars means every open in the window was exactly its low, which does not force Σ(high − open) to zero — a market that opened on its low every day and rallied every day is a real (if extraordinary) tape, and its A ratio is x/0. Applying the package's flat-window test (#699: is the numerator forced to zero, and is the answer determined?) both answers are no, so the reading is missing. The guard is live, and without it the value would be ±Infinity, which withColumn rejects outright.
• A negative sum is possible only by redirecting a column (an open pointed at a smoothed price can sit outside the bar), and reads honestly: a negative ratio means the two sums disagreed in sign. Nothing clamps it.