Every number here, and how it was worked out

A day of trading gives us four prices: the open, the high, the low and the close. We also keep yesterday’s close, high and low, because most of what matters is a comparison between today and yesterday.

The return

The headline number. How much did the market move, close to close?

return = (close − previous close) / previous close

The gap

Did today open clear of yesterday’s entire range? Not merely above yesterday’s close — above yesterday’s high.

gap up   ⟺  open > previous high
gap down ⟺  open < previous low
otherwise, no gap

On our dataset, 16.73% of days gap up and 7.92% gap down. Those do not sum to 100% — the remaining 75% opened inside yesterday’s range. This matters later: gap up and gap down are two separate questions, not two halves of one.

The intraday move and the day’s range

intraday = (close − open) / open
range    = (high − low) / open

Take an axis — say Moon’s nakshatra. Every trading day in thirty years falls into exactly one of its 27 values. That gives 27 buckets, and each bucket is simply a list of dates. “Moon in Ashwini” might be 280 days scattered across the decades.

Four families of axis exist, and they differ in what a bucket is:

Two rules that shape every bucket

A day only counts if it has a previous close. The very first bar of the price history has nothing to compare against, so it forms no return, no gap, no outcome — and it is dropped. Every metric in a bucket therefore shares one denominator.

An event is mapped to a trading day. A conjunction that becomes exact on a Sunday did not happen on a trading day, so it is carried to the next session. This is why an event bucket’s day count can differ from its raw event count — and why the app says so explicitly rather than hiding the difference.

raw Moon–Jupiter conjunctions found       545
  −  before NIFTY price history begins       78
  −  dated in the future                     59
  −  landing on the very first price bar       1
                                          ─────
  usable occurrences                       408

Win rate

win rate = up days / (up days + down days)

Across all 7,605 days, NIFTY’s unconditional win rate is 0.5327 — it closes up 53.27% of the time. That single number is the yardstick every single-axis bucket is measured against, and it is worth pausing on: the market goes up more often than it goes down. Any bucket that wins 53% of the time is unremarkable, not bullish.

Bias

bias = (up days − down days) / (up days + down days)

Average return — and why it can disagree

average return = (sum of all daily returns) / n

Win rate counts how often. Average return measures how far. They answer different questions and can point opposite ways: a bucket that rises on 60% of days by a tiny amount and falls on 40% of days heavily has a good win rate and a negative average return. Neither number is wrong; they are measuring different things. This is precisely why the app later lets you choose which one the significance test is run on.

Volatility

                  ______________________
volatility  =  √ Σ(return − mean)² / (n − 1)

Volatility has no good side. A combination that reliably doubles volatility is genuinely notable and genuinely tradeable, but it is neither bullish nor bearish — which is why the app colours it by size rather than by direction.

Forward returns

forward 3-day return = close[day + 3] / close[day] − 1

This asks whether a condition leaves a mark that outlasts the session it appeared in. It also introduces a subtle problem that Part 10 has to fix.

If you flip a fair coin ten times you will not get exactly five heads. Six is common, seven happens. So when a bucket of ten days shows seven up-days, “70%” is not the truth about that bucket — it is one noisy sample of it. A confidence interval is a range that says: given this sample, the underlying rate is plausibly somewhere in here.

The app uses the Wilson score interval at 95%. The obvious alternative — the “normal approximation” most textbooks teach first — breaks badly on small or lopsided samples, happily producing ranges that extend below 0% or above 100%. Wilson cannot: it is constructed so the interval always lies inside the possible range.

p = k / n            (the observed rate)
z = 1.96             (for 95% confidence)

              p + z²/2n            z          ⎧  p(1−p)     z²   ⎫
centre  =  ─────────────    half = ───── × √ ⎨ ─────── + ───── ⎬
             1 + z²/n             1 + z²/n    ⎩    n       4n²   ⎭

interval = centre ± half

p  = 24 / 40                         = 0.600
z² = 1.96²                           = 3.8416
z²/n = 3.8416 / 40                   = 0.0960
1 + z²/n                             = 1.0960

centre = (0.600 + 0.0480) / 1.0960   = 0.591
half   = (1.96 / 1.0960) × √(0.006 + 0.0006)
       = 1.788 × 0.0813               = 0.145

95% interval                          = 0.446 to 0.736

This is why the app shows n beside every rate, and why the drill-down reports the interval rather than the bare percentage. The width of the interval is the honesty of the number.

Suppose a bucket contains 300 trading days. That sounds like a lot. But if those 300 days are one unbroken stretch — say, the period Pluto spent in a single nakshatra — then you have not observed the condition 300 times. You have observed it once, for a long time, and whatever else was happening in the market during that year is baked into every one of those days.

So the app counts spans: the number of separate episodes.

a new span begins when
    the bucket value changes
  OR the previous matched day is not the immediately preceding trading day

Neptune in Pisces (Meena)
    days  955     spans   2     average run   477 days

Pluto in Purva Ashadha
    days 1662     spans   5     average run   332 days

Moon in a given nakshatra
    days ~280     spans ~280    average run   ~1 day

The app uses spans in three places. Buckets need at least 8 spans as well as 30 days before they are counted as testable. The sample-quality badge (thin / moderate / strong) reads spans alongside n. And the KP explorer prints a warning above any axis whose average run exceeds 30 days, calling those buckets eras, not recurring conditions.

A bucket wins 58% of the time where the market wins 53%. Is that something? It depends entirely on how many days are behind it. Five percentage points on 30 days is noise. Five points on 3,000 days is not. We need one number that folds both together.

That number is the Z-score, and it works by asking: if this bucket were really no different from the baseline, how far from the baseline would a sample this size normally wander? That typical wandering distance is the standard error. Z is simply how many standard errors away we actually landed.

                observed − expected
Z  =  ───────────────────────
              standard error

                    _____________
standard error = √ p(1 − p) / n           (for a rate)

expected p = 0.5327   (NIFTY's own baseline)
observed   = 0.58

n = 30      SE = √(0.5327 × 0.4673 / 30)   = 0.0911
            Z  = (0.58 − 0.5327) / 0.0911  = 0.52

n = 3000    SE = √(0.5327 × 0.4673 / 3000) = 0.0091
            Z  = (0.58 − 0.5327) / 0.0091  = 5.19

Reading a Z-score

Note the shape of that table. Z is not linear — the difference between 3 and 4 is a factor of forty. It is also signed: a Z of −3.2 is exactly as notable as +3.2, just in the other direction. The app sorts on the magnitude for that reason.

Suppose Moon in Chitra wins 57% of the time, and Venus in Kanya wins 56%. What should the combination win? If the two conditions simply stack with no interaction, the combination should land somewhere near — but where exactly?

Not their product. 0.57 × 0.56 = 0.32, and a 32% expected win rate is obviously absurd. Multiplying is right for the probability of two independent events both happening; it is wrong for combining two rates that each already describe the same outcome.

The app uses a log-odds model instead. Odds are just a rate re-expressed: a 57% win rate is odds of 0.57/0.43 = 1.33 to 1. Taking logarithms turns “combining” into ordinary addition, which is exactly the behaviour we want.

logit(p) = ln( p / (1 − p) )

logit(expected) = logit(row) + logit(column) − logit(overall)

expected = 1 / (1 + e^−logit(expected))

row (Moon)      0.57    logit = ln(0.57/0.43) = +0.2819
column (Venus)  0.56    logit = ln(0.56/0.44) = +0.2412
overall         0.5327  logit = ln(0.5327/0.4673) = +0.1310

logit(expected) = 0.2819 + 0.2412 − 0.1310  = 0.3921
expected        = 1 / (1 + e^−0.3921)        = 0.5968

This is why the heatmap is coloured by residual and not by raw win rate. A grid coloured by win rate mostly shows you which rows and columns are individually strong — information you already have from the margins. Colouring by residual shows only what the combination contributes, and a cell that is merely “hot because one of its axes is hot” correctly reads as neutral.

Two different nulls, two different questions

A dashboard row therefore carries up to two Z-scores, and they answer different questions:

A combination can beat the market handsomely and still contribute nothing — because both of its parts already did. The vs market column is shown only when it is not simply a repeat of Z.

A |Z| of 3 happens by chance about once in 370 tests. That sounds reassuringly rare — until you notice the app tests 296,914 combinations for one market. At one in 370, pure chance should hand us roughly 800 cells at |Z| ≥ 3, every one of them meaningless.

Cricket makes the shape of the problem easy to see. Slice thirty years of scorecards finely enough — this batsman, at this ground, batting second, in April — and one of those slices will show a staggering average. It is staggering because it rests on a handful of innings, and you found it precisely because it was the most staggering of the thousands of slices you cut. Singling out the best-looking of our 296,914 combinations and calling it a discovery would be the same act: reporting the slice, not the cricketer. It is an easy mistake to make, and it is how a great many market patterns — planetary and otherwise — end up published.

The Bonferroni correction

The fix is to raise the bar in proportion to how many tests you ran. If you want a 5% chance of any false positive across the whole search, each individual test must clear a much higher threshold.

bar = the |Z| for which the chance of a single false positive
      is  α / (number of tests)        with α = 0.05

a cell at |Z| = 4.33
    in a single test        → about 1 in 75,000 by chance   → striking
    among 296,914 tests     → about 4 such cells expected   → unremarkable

the bar for 296,914 tests                                   |Z| ≥ 5.23
the largest deviation actually found anywhere               |Z| = 4.33

The second disclosure: found versus expected

The app also reports a looser comparison that is easier to check by eye. Of the 292,513 two-axis combinations large enough to measure, how many reach |Z| ≥ 2 — and how many should, by chance alone?

cells large enough to test                     292,513
chance rate of |Z| ≥ 2                          ≈ 1 in 22
expected by chance alone                       ~13,309
actually found (direction)                      7,569

Until recently every Z-score in the app measured exactly one thing: how often the market closed up. Volatility, gap frequency and the return columns were all displayed, but nothing tested them. A combination that reliably doubled volatility, or reliably gapped down while closing flat, was statistically invisible.

You can now choose which quantity the significance test runs on. The rows, the day counts and every descriptive column stay exactly the same — only the Z changes.

Three kinds, three formulas

The kind is not a label — it determines both the test statistic and the way two conditions are combined.

proportion   Z = (p − E) / √( E(1−E) / n )
             combine by log-odds        (Part 7)

mean         Z = (m − E) / ( sd / √n )
             combine by addition:  row + column − overall

dispersion   Z = ( ln s − ln E ) / SE
             combine multiplicatively: row × column / overall

The two gap factors are deliberately kept separate rather than combined into one. They are not mirror images: a day that gaps neither way counts against both denominators, so their rates do not sum to 1 (16.73% and 7.92% leaves 75% of days gapping neither way) and each carries its own information.

Directionality drives the colour. Direction and the two return factors are signed — up is bullish, and the heatmap uses the familiar green-red. The gap factors and volatility are not: a busier market is notable but it is not good, so those get a single amber magnitude scale instead. Colouring a high-volatility cell green would be a claim we do not mean to make.

The standard error formulas in Part 6 all rest on one assumption: that each observation is an independent piece of evidence. For “did the market close up”, on daily data, that is roughly true. For three of our six factors it is badly false, in three different ways.

Correction 1 · Overlapping windows

The 3-day forward return has a structural problem. Monday’s three-day window covers Tuesday, Wednesday and Thursday. Tuesday’s covers Wednesday, Thursday and Friday. They share two of their three days. Consecutive matched days are therefore measuring almost the same stretch of market — but the formula counts them as two independent observations.

correlation between two windows d days apart  =  max(0, 1 − d/h)

                    n²
effective n  =  ─────────────
                 Σᵢⱼ corr(i,j)

60 days, all consecutive (one unbroken run)
    Σ corr ≈ 60 × 3        effective n ≈ 20   →  Z shrinks by √3 = 1.73×

60 days, all more than 3 trading days apart
    Σ corr  = 60           effective n  = 60   →  Z unchanged

Correction 2 · Clustering

Gaps are not spread evenly through history. They arrive in bursts when markets are turbulent and go quiet for months when they are calm. And the gap test is doubly exposed, because its threshold is yesterday’s range — so gap frequency is mechanically tied to how volatile the market currently is.

We can measure how badly this breaks the independence assumption directly. Chop the history into blocks of 20 trading days, count the gaps in each block, and compare how much those counts actually vary against how much a coin-flip model says they should.

                observed variance of block counts
inflation (VIF) = ───────────────────────────────────
                     n × p × (1 − p)     [the binomial prediction]

corrected n = n / VIF

factor        lag-1 correlation    VIF     effect on Z
direction          +0.075          1.19    negligible
gap down           +0.078          2.41    ÷ 1.55
gap up             +0.137          3.43    ÷ 1.85

before correction   gap up reached  |Z| = 6.22   (bar 5.23 — "a finding")
after correction    gap up reaches  |Z| = 3.36   (comfortably under)

Correction 3 · Non-stationarity

The third problem cannot be fixed by adjusting a sample size, and understanding why is worth the paragraph.

Market volatility comes in regimes that last for years. Now recall Part 5: a bucket on a slow axis is a contiguous era. Neptune in Pisces is 955 trading days in two spans. Comparing that bucket’s volatility against a thirty-year average — which pools calm decades and crisis years together — does not measure Neptune. It measures whichever years Neptune happened to be passing through.

measured inflation for volatility, by block length:
    30 days    VIF  2.3
    60 days    VIF  4.1
   120 days    VIF  7.2
   250 days    VIF 13.4

A well-behaved quantity converges to a stable VIF.
This one grows without settling — the signature of a
process whose memory never really fades.

The app’s answer is to withdraw that specific reading rather than the whole factor. Volatility keeps its Z wherever the comparison is drawn from the same set of days — heatmap cells, scanner cells, and two-axis combinations, all of which compare a cell against margins built from exactly the days being measured, so the regime cancels out. Only the comparison against the thirty-year baseline is dropped, and those cells show “—”.

Every correction so far makes a single measurement more honest. None of them answers the question that actually matters to someone thinking about money: does it keep happening?

The scanner answers that by splitting history in two — the first 70% to search in, the last 30% held back. The split is strictly chronological, never random. A random split would scatter days from the same week into both halves, and since neighbouring days share market conditions, the held-back set would already contain most of the answer.

One detail carries a lot of weight: the holdout’s expectation is rebuilt from the holdout period’s own row and column rates, never from the training period’s. Reusing the training rates would leak the very bias the check exists to detect, and the consistency flag would become decorative.

Excluding pairs that cannot be independent

Some axis pairs are guaranteed to agree, and testing them would be meaningless. Two layers of exclusion apply:

• Structural — one axis is a direct lookup of the other. A planet’s sign determines its sign-lord completely, so pairing them tests nothing. Rahu and Ketu are always exactly 180° apart, which makes their signs a lookup — though notably not their nakshatras, since 180° divides evenly into 30° signs but not into 13°20′ nakshatras. Twenty-one pairs are excluded this way.
• Measured — a symmetric uncertainty coefficient, computed on the actual days in your current selection, drops any pair that turns out near-redundant in practice.

Worth noting what is not excluded: a planet’s star-lord paired with its sub-lord — the single most-wanted KP combination — survives both gates, because a star lord genuinely admits nine different sub lords. It is a real pairing, not a lookup.

Everything so far measured what the market did. The backtest asks a harder question: if you had acted on this, what would have happened? That requires deciding when you get in, when you get out, and at what price — and each of those decisions can flatter the result if made carelessly.

Getting in

Breakout entry is the more honest of the two in one respect: you are not choosing a direction in advance, the market is telling you. It also means the direction setting is ignored in that mode, which the interface reflects rather than leaving as a trap.

Getting out

Four exits compete, and the first one hit wins: a stop loss, a trailing stop, a take profit, or simply running out of holding days. The simulation walks forward one bar at a time using the full high-low range, not just closes.

within a single bar, the conservative order is assumed:
    stop is checked BEFORE target

if the bar opens beyond the level (a gap through):
    the fill is the OPEN, not the level

long entry at            100.00
stop placed at            99.00   (a 1% stop)

next day the market opens at 93.00 — the stop level
was never traded; the market gapped straight past it.

fill  = 93.00  (the open, not 99.00)
loss  = −7.0%  on a "1% stop"

Scoring the trades

Returns are sign-adjusted: a short that profits from a falling market is recorded as a gain. Beyond the average and the win rate, two numbers matter:

max drawdown = the largest peak-to-trough fall
               of the running equity curve,
               with trades taken in date order

Here is the complete result for NIFTY’s closing session across 7,605 trading days, measured through 7 August 2026. Each factor is tested independently, and each gets its own bar because each searches a slightly different population.

Nothing has cleared the bar on any factor yet. The largest deviation found anywhere, across all six ways of measuring and nearly three hundred thousand combinations, is |Z| = 4.75 — against a bar of 5.23. And on every single factor, the number of combinations reaching the looser |Z| ≥ 2 threshold is below what chance alone predicts.

Volatility’s population is smaller (292,513) because, as Part 10 explained, it performs no test against the thirty-year baseline at all.

Every number in that table is NIFTY 50, closing session. The other 223 markets the app carries — 200 individual stocks and 23 further indexes, most with a decade or more of history — are each waiting for the same search, and it is the same three clicks to run. So this is a result about one index measured one way, not about the question.

Where this leaves the search

This is where the evidence stands today, and it is worth reading precisely rather than as a verdict:

• It does not show that planetary positions have no relationship to prices. Absence of evidence at this sample size and this correction level is not evidence of absence.
• It does narrow where an effect, if there is one, could still be hiding. A large, simple effect that held steady for thirty years — a nakshatra that reliably moved the Indian index by a tradeable amount — is exactly the kind of thing three hundred thousand tests are well positioned to catch, and these axes have not caught it on this index. What is left unsearched is where the search goes next: the 223 other markets in the app, finer conditions, effects that come and go with the era rather than holding for three decades — and, above all, the possibility that these conditions do something together that none of them does alone — which is Part 15, and the one you can go and check today. Part 14 sets the limits out one by one and Part 16 is the practical version; each of them is a direction rather than a dead end.
• It leaves the base rates genuinely useful in the meantime. Knowing that a condition historically closed up 56% of the time, with a confidence interval wide enough to be honest about it, is real context for a chart you are already reading. It is not a signal, and the app never presents it as one.

There is something to take from the shape of the null, too. On every factor, fewer combinations reach |Z| ≥ 2 than chance alone would produce. A method with a thumb on the scale — a correction dropped here, a bar set conveniently there — shows the opposite symptom: a surplus of near-misses, not a shortfall. Whatever this method eventually finds will be worth believing for the same reason it has found nothing yet: the corrections in Part 10 are applied whether or not they are convenient, the bar is set by the size of the search rather than by the result, and the banner on the dashboard reads the live numbers instead of a hard-coded sentence — so it stops saying “nothing clears it” on the day something does.

One market measured, many more not

Every figure here is NIFTY 50. A pattern absent from one index may well exist in another, and a pattern present in one market becomes far more convincing the moment it replicates in a second. The app holds 200 stocks and 24 indexes, so that argument is available to be made — it simply has not been made yet. It is one of the two biggest things left to go and find out — the other is Part 15’s — and Part 16 is about how to start on it.

Thirty years is fewer observations than it sounds

7,605 trading days is a lot of rows and not much independent evidence for anything slow. Saturn completes roughly one orbit. Any statement about a body that changes sign every few years rests on a handful of episodes, whatever the day count says — which is exactly what Part 5’s span counting exists to surface.

Two ayanamsas that must not be mixed

The KP screens use the Krishnamurti ayanamsa; the Daily and Panchang screens use True Chitrapaksha. Measured against each other, they sit 0.075°–0.087° apart in longitude and agree on the star-lord 99.44% of the time — but they are different frames, and a chart that mixed a condition from each would describe a sky that exists in neither system. The app therefore refuses to combine them rather than offering the option with a warning.

A deliberate omission

Karana is not offered as an axis. Eight of its eleven values are stored with a broken cycle in the underlying data, and the same defect appears in the event tables — an upstream generation bug, not a display issue. Rather than present a broken axis with a caveat, it is left out until the data is fixed.

Counts that legitimately disagree

A backtest’s occurrence count need not equal the day count of the bucket it came from. The backtest counts each event date once and drops events with no forward price data; the breakdown counts trading days and drops the first bar. Same underlying events, different tally, and the app says so on screen rather than letting you find the discrepancy yourself.

Survivorship and revision

Index history is not a neutral record. NIFTY’s constituents have changed many times; the index you can measure today is not the portfolio anyone held in 1996. This affects long-run return figures more than it affects direction or gap frequency, but it affects everything to some degree.

Look at the grain Part 13’s numbers are measured at. Each of those 296,914 tests asks about one thing: a single planetary event, a single Daily position, a single KP lord — or, in the two-axis rows, a screened pair of them. Precise, and deliberately narrow. It is the grain a machine can search exhaustively.

A day you would actually trade is not one condition. Open the Dashboard on any date and it lists every planetary condition live on it, each carrying its own thirty-year base rate — the whole configuration, in one screen, the way you would read a chart. Measured on NIFTY’s closing session across twenty sampled trading days, a single day carries about 424 measured combinations — roughly 360 KP, 44 Daily, 11 planetary events and 4 Panchang, though the exact mix changes from day to day.

Which means Part 13’s headline is narrower than it first sounds. “Nothing has cleared the bar” describes conditions taken one and two at a time. Read at that grain, a day where a dozen conditions all lean the same way and a day where they pull against each other look identical. Read whole, they are obviously different days — and that difference is the thing you can go and put to the test.

And it is the easier question to answer

Here is the part that makes this genuinely promising rather than merely appealing. Part 8’s bar is high — |Z| ≥ 5.23 — for one reason: 296,914 tests were run, so the bar had to rise to match. Widening that search makes the problem worse, not better. Triples on the KP axes alone would multiply it by orders of magnitude and push the bar further out of reach.

A whole-day reading is the opposite kind of question. It is one hypothesis, not a search: does a day’s configuration, summarised into a single number, tell you anything about that day? One question faces a bar of roughly |Z| ≥ 2, not 5.23 — a threshold thirty years of data can comfortably reach. The atomic search is the hard way to find something. This is the tractable one, and it is sitting right there on the Dashboard.

What that summary should be is the genuinely interesting part, and it is wide open: a count of the conditions that agree, an average of their biases, a weighting by sample quality, by span count, or by family. Pick one, look at enough days, and you have a result nobody has.

Go and check a day

You need no tooling you do not already have — the reading is the skill you brought with you. Open the Dashboard (it takes a free account), pick a date, and read the day whole rather than row by row. Which conditions are live together? Do they agree? Then scrub to the next day and the one after, and watch whether the days where everything lines up behaved differently from the days where it did not.

Then tell us what you found. These are the answers we most want: do some combinations carry more weight than others, and if so which — the rare ones, the ones with the most spans, the ones from a particular family? Does a day with many agreeing conditions behave differently from a day with few? Is there a handful that matters and a long tail that is noise? Say it with the dates you looked at, so it can be checked. The Community board is the place — a finding posted there is visible to everyone who comes after, which is what turns one person’s observation into something worth testing properly.

The Explorer, the Scanner and the Backtest all begin with a market selector, and it is not a short list. The app carries daily price history for 200 individual stocks and 24 indexes — NIFTY 50 and SENSEX, the Bank, IT, Auto, Pharma, FMCG, Metal, Energy, Realty, Media, PSU Bank, Infrastructure, MNC and Consumption sector indexes, and the Midcap and Smallcap families. 173 of those 224 carry ten years of history or more; the deepest runs to 8,986 trading days.

Part 13’s table covers exactly one of them. NIFTY 50 is the market the significance dashboard is currently pinned to, and it is where the corrections in Part 10 were measured — but the search it describes stops there. The other 223 searches are sitting behind a dropdown, unrun.

Things worth trying

• Read a whole day, not a row. Part 15 leads this list because it is the one you could answer first. Every other item below refines a question the app already answers well. That one asks something new of a screen you already have, and it faces a far lower bar for the reason Part 15 sets out.
• Take a NIFTY pattern and check it on a sector index. A combination that leans the same way in NIFTY Pharma and in NIFTY Metal — two indexes with almost nothing in common — is worth more than a larger deviation in NIFTY alone. Replication across unrelated markets is the strongest currency this field has, and no amount of further testing on a single index can buy it.
• Look at single stocks, not only indexes. An index is dozens of companies averaged together, and averaging is precisely what dissolves an effect that lives in one sector or one kind of business. If a planetary condition means something for a metals company, an index where metals are a small slice will dilute it toward invisibility.
• Read spans before you read day counts. Part 5 is the trap most searches fall into. A bucket of 300 days in 2 spans is one observation wearing a large number, and the fast bodies — Moon above all — are where a big n is genuinely made of hundreds of separate episodes.
• Change the factor before you change the axis. Part 9’s six factors ask six genuinely different questions of the same days. A condition that does nothing to direction may do something to volatility or to the gap, and until recently nothing in the app tested those at all — which means they are the least picked-over ground here.
• Try both sessions. The 09:15 and 15:30 snapshots disagree on the tithi for 26.5% of dates. That disagreement is not noise to be averaged away; it is an open question about which moment of the day the sky is supposed to be read for.
• Carry it through to the backtest. A base rate is not a strategy. Part 12 exists so that a bucket you find interesting becomes an explicit rule — entry, stop, exit — and gets measured as a trade rather than admired as a percentage.

Tell us what is missing

This app expresses a particular family of questions: a position, a lord, an almanac value or an event, taken one or two at a time and measured against a trading day. That is not the whole of what practitioners actually use. The gaps are not hidden by design — they are mostly the things nobody has yet asked for in a form precise enough to build.

So the most useful thing you can send is a technique this app cannot currently express, stated the way you would state it to another practitioner: which bodies, which condition, read at which moment, and what you expect it to do to the price. A question phrased that precisely is most of the work of building it. Half-formed ones are welcome too — more than one thing now in the app began as somebody asking why a screen did not offer something obvious.

Two places to do that: the Community board, where the question and whatever comes of it stay visible to everyone, and Contact, which needs no account, if you would rather it came straight to us. Corrections to this page belong there too — if a formula here is wrong, or a caveat is missing, saying so is a contribution.

Two screens carry all of this. The Explorer is where a market, an axis and a bucket become the numbers Parts 3 to 7 describe. The Dashboard is where a whole day arrives at once — and that is the one to open first, because the question it raises is the one nobody has answered yet. Pick a day, read what is active on it together, and tell us what you see.

Reference values · NIFTY 50 closing, 7,605 days through 2026-08-07

Every number on this page was measured on the live dataset rather than illustrated. Where a figure would change as new price data arrives, it is quoted with the date it was measured. The methods described here are implemented in the application itself and covered by its automated tests — including tests whose specific purpose is to fail if any factor ever does clear the significance bar, so that a real finding would be reported rather than missed.