What you can build with this data
Nine years of who is levered into what, stock by stock, every day. Filter the ideas below, click any card to see how to build it. Methodology only — no views on individual stocks.
9 years
daily history since 2017
2,400+
funded stocks, NSE + BSE
~1,700
non-F&O names with no other positioning data
4 fields
₹ book · shares · implied price · close
Evidence — what testing already showed
Backtested on 1,741 non-F&O NSE stocks, 2017–2026. A measured edge is not a promise.
Crowding hurts — and so does panic
Excess return at 63 days, by how the book moved
Spikes = volatility, not direction
63-day outcome spread after a book spike (13,700 events)
Context decides the sign
Excess return at 21 days, by what preceded the spike
Is this spike worth anything?
Run any book jump down these four questions.
Did shares rise?
If only ₹ rose, the price moved — not the financing. Require Δqty > 10%.
else ignore
Is it big enough?
≥ 25% and ≥ ₹1 Cr, so tiny books don't dominate.
else ignore
Had it already run?
Up >10% over the prior 20 sessions — the chase bucket, ~40% of spikes.
fade −0.8 to −1.1pp
Or no prior move?
Flat price, book explodes. Monster subset ran +0.6pp @21d (n≈450).
watch
Ideas
Sorted by how ready they are to build. Click to expand.
What it says
How to build it
Before you backtest anything
Six things that will quietly invalidate results.
Publication lagOne Friday book only arrived the following Tuesday. Using it on its disclosure date is look-ahead. Lag by worst realistic delay.
Splits & bonusesShare counts jump for non-events. Require the ₹ book to move too.
Float is a snapshotToday's float, applied to 2018, imports look-ahead. Restrict the lookback or source historical float.
SurvivorshipDelisted names are in the mirror. Include them or long-horizon results flatter.
LiquidityThe strongest signals sit in micro caps. Model impact before believing the edge.
No shortingNon-F&O names can't be shorted. The contrarian leg is a veto, not a short book.
Start here. Two cheap experiments with real information value: confirm what the implied financed price actually is (one query — it either opens a new signal or closes it), and re-run the buildup factor with a proper publication lag, which decides whether the measured edge survives into something implementable.