How it learns

A system that is judged by its own record, and re-tested against it.

Every call is checked against Nifty after ten sessions and kept, hits and misses alike. At least once a month the scoring weights are re-fitted on that record and tested on seasons the fit never saw. They change only when the evidence clears a bar set high enough to rule out luck. Everything on this page is the stored record, drawn as it stands.

856: shares read from the close of 6 Oct
3,630: named calls judged · 50 live, the rest replayed
8: re-tests on record · 4 passed and went live
1 Nov: the next scheduled re-test

The loop

it runs every trading day; the last step once a month

1Read856 shares
After each close, six layers score every liquid share on the exchange. The top five and bottom five are named.
2Wait10 sessions
Nothing is judged early. Each named call runs ten sessions, watched close by close.
3Judge3,630 judged
Did a look-at name beat Nifty, and a stay-away name trail it? Both answers are kept.
4Record283,067 in the record
Every judged read joins the record. The chance beside each name is its rank band's hit rate, refreshed daily.
5Re-testMonthly · 1 Nov
Weights are re-fitted on past seasons and scored on seasons the fit never saw. Kept only if they clear the bar.

…and the next evening's read runs on whichever weights won.

The record it learns from

named calls, judged ten sessions on against Nifty

52% of 3,630 named calls went the way they leaned. A coin would sit at 50%.

live monthsreplayed months (the read run on past data, labelled simulated)share that went the way it leaned50%, a coin

Every re-test, kept or rejected

fitted on earlier seasons, scored only on later ones

The daily read · today's five, from every close

To go live, new weights need a t-statistic of 3.5 or more over at least 120 sessions, measured out of sample, and must beat the starting weights.

Fitted	Method	Out-of-sample IC	t	Mornings	Outcome
1 Oct	ic	0.039	10.2	300	Passed · live ic: out-of-sample IC +0.0394 (t 10.17) over 300 mornings vs prd +0.0239
19 Sep	ic	0.042	9.5	292	Rejected ic: out-of-sample IC +0.0421 (t 9.47) over 292 mornings vs prd +0.0225
15 Sep	ic	0.040	10.0	289	Rejected ic: out-of-sample IC +0.0401 (t 10.04) over 289 mornings vs prd +0.0227

The results read · companies about to report

To go live, new weights need a t-statistic of 2 or more over at least 20 mornings, measured out of sample, and must beat the starting weights.

Fitted	Method	Out-of-sample IC	t	Mornings	Outcome
1 Oct	ic	0.086	4.5	28	Passed · live ic: out-of-sample IC +0.0860 (t 4.51) over 28 mornings vs prd +0.0130
19 Sep	ic	0.094	4.6	28	Rejected ic: out-of-sample IC +0.0942 (t 4.59) over 28 mornings vs prd +0.0130
15 Sep	ic	0.088	4.6	28	Passed · live ic: out-of-sample IC +0.0876 (t 4.56) over 28 mornings vs prd +0.0139
15 Sep	ic	0.097	5.1	28	Passed · live ic: out-of-sample IC +0.0969 (t 5.06) over 28 mornings vs prd +0.0139
15 Sep	ic	0.097	5.1	28	Rejected ic: out-of-sample IC +0.0969 (t 5.06) over 28 mornings vs prd +0.0139

The weights, then and now

how much each layer counts in the score

The daily read

Running on weights from the re-test of 1 Oct.

Reactivityhow far it has moved on its own results 0%as started
Momentumits move over 30 and 90 sessions and the 52-week range 18%−22 pts
Divergencewhether it has pulled away from Nifty 0%−35 pts
Sectorits strength against its own sector index 0%−25 pts
Own recordhow it has done after its past results 60%+60 pts
Last quarterwhat its latest reported quarter showed 22%+22 pts

The results read

Running on weights from the re-test of 1 Oct.

Reactivityhow far it has moved on its own results 0%−30 pts
Momentumits move over 30 and 90 sessions and the 52-week range 15%−5 pts
Divergencewhether it has pulled away from Nifty 0%−30 pts
Sectorits strength against its own sector index 0%−20 pts
Own recordhow it has done after its past results 85%+85 pts
Last quarterwhat its latest reported quarter showed 0%as started

weight in forcestarting weight A layer that starts at zero can only earn weight through a passed re-test.

Where the "chance" figure comes from

each tenth of the field, its own record

Every name read is bucketed by where it ranked that day, and each tenth's record is kept. The chance printed beside a name is simply its tenth's hit rate here — learned from the record, refreshed every day, and hidden until a tenth holds 100 judged reads. A ranking that carries information shows a staircase; a flat row would mean it does not.

51% Top tenth 28,426
49% D2 28,248
48% D3 28,292
48% D4 28,253
48% D5 28,225
46% D6 28,324
46% D7 28,288
47% D8 28,257
47% D9 28,283
48% Bottom 28,471

Bars: share that beat Nifty over the next ten sessions. Small number: reads in that tenth. "…" = under 100 judged, not yet shown on cards.

Learning indexout of sample · 4,419 judged reads

+0.086

information coefficient of the model in force · t 4.5 · positive on 89% of 28 mornings

25Q3 25Q4 26Q1 26Q2 26Q3

hypothesis weights on the same mornings: +0.013 · last fit 1 Oct, applied

Profitability indextop fifth vs bottom fifth · ten sessions · vs Nifty · before costs

+28.7 pts

cumulative long-short excess over 29 result weeks · +17.7 pts after 20 bp a position · +1.18% per read · top fifth beat Nifty 42% of 886

hypothesis weights, same procedure: +20.0 pts · a description of the record, not a recommendation

What never learns

fixed by design, so the record stays comparable

The layer formulas. What each layer measures is fixed; only how much each counts is re-tested.
The cut-offs. Five named each side on the daily read, the top and bottom fifth on the results read, ten sessions, Nifty as the yardstick, a 20 basis-point cost per position in the indices.
The bar itself. The test a re-fit must pass is fixed in advance and never loosened to let a fit through.
Market regimes. Used to break the record down, never to switch weights.
The past. A read is never rewritten. Replayed months stay labelled as simulated.

The full numbers are on the track record; the working of every layer is in the methodology. Our calculation, not advice.