Delivery % is not smart money: a test on 173,985 volume spikes
NSE delivery data since 2010, BSE since 2024. We tested high against low delivery on volume spikes. The gap is real, but it mostly sits where you cannot act on it, and the popular screen trailed its own universe.
Indian stock screeners show a delivery % column, and the usual reading is that high delivery means strong hands are accumulating. NSE's own header calls the number Deliverable Quantity(gross across client level). It says nothing about who bought, or for how long. A fund buying for five years and a trader holding until Thursday look the same.
We now publish delivery data free: NSE from 4 January 2010, BSE from 8 July 2024. So we tested the story, and the answer is more awkward than "it's noise". This post covers the data, two things we found wrong while writing it, and the test.
Two files that do not look alike
NSE's is the security-wise delivery position file, MTO_DDMMYYYY.DAT, out around 16:30 to 17:00 IST. BSE's is gross deliverables, SCBSEALLDDMM.zip, around 18:15 to 20:30 IST. BSE's file names no symbol, only a scrip code, which is the same day's bhavcopy's FinInstrmId, so the bhavcopy names the rows. The regenerated 25 April 2024 file spells its header DAY`S VOLUME, so we read columns by position.
Across 579 BSE days and 1,352,226 rows there are zero volume mismatches, zero duplicates, and delivery_pct agrees with quantity over volume. Seven bse/ rows have no delivery row: trades of 2 to 10 shares in odd symbols like SBIN# and IDEA#.
What we got wrong
Our pipeline's notes say NSE delivery volume equals bhavcopy volume. Checking that for this post, it holds on 6,691,008 of 6,696,836 joined rows, or 99.91%. The other 5,828 are on six dates, most of them on three:
date rows off RELIANCE, bhavcopy RELIANCE, MTO file
2026-09-11 2,584 8,777,736 7,451,218
2019-06-18 1,244
2019-06-17 1,236 6,815,554 7,042,630
2025-10-21 748
2013-05-13 14
2023-09-04 2We downloaded the raw files for 11 September 2026 and 17 and 18 June 2019, and all 5,734 EQ rows in them match what we publish, so the numbers are NSE's, not ours. On those dates delivery_pct is against the MTO file's own volume, so it will not equal quantity over bhavcopy volume. Separately, the files for 17 and 18 June 2019 list 123 rows under numeric codes like 500068 where a symbol should be. We do not know why. We left all of it as published and are telling you instead.
The second problem is ours. Our point-in-time universe, universe/nse_liquid, has no rebalance for July, August or September 2011: it goes from June straight to October. We found it because a shorter version of this test disagreed with the long one by 2,952 events. It is not fixed yet. Everything below skips those 63 sessions.
Delivery over time, and by size
NSE is EQ series only, and that includes ETFs. The number depends heavily on how you weight it:
all NSE EQ value-weighted, by turnover rank (from May 2010)
year value-wtd median row 1-100 101-250 251-500 rest
2010 39.90 56.84 41.26 40.98 40.51 34.11
2013 45.75 68.06 45.44 44.94 52.98 52.15
2015 46.63 61.02 47.88 41.60 47.64 46.48
2020 31.44 59.38 29.39 35.02 45.80 46.08
2022 41.10 53.33 42.28 37.75 39.72 42.97
2025 44.82 53.49 48.76 43.30 39.31 39.74
2026 42.85 54.24 47.20 42.15 38.42 35.78Two things stand out. The typical row has become less delivery-heavy, from 68.06 in 2013 to 52.41 in 2024, the lowest year. Meanwhile value-weighted delivery fell to 31.44 in 2020 and recovered. And the size relationship flipped: in 2020 the top 100 by turnover were well below the rest of the market, and in 2025 they were well above it. These files cannot tell us why, so we will not guess.
The other trap is thin trading. Over the whole sample the median row outside the top 500 is 62.45% against 40.25% in the top 100, but value-weighted they are 40.90% and 41.83%. When few shares trade, a high percentage is easy: 167,239 NSE rows are exactly 100%. A high number on a small stock may mean little more than that.
The same stock, two exchanges
Since 8 July 2024, value-weighted delivery is 54.17% on BSE and 43.78% on NSE. Pair the same ISIN on the same day, 1,139,219 pairs across 2,828 ISINs, and BSE is higher 70.1% of the time with a median gap of 8.38 points. The correlation is only 0.57. Limit it to pairs with at least ₹5 crore on both exchanges and the gap barely closes: 7.06 points, correlation 0.55. On 4 November 2026, RELIANCE was 50.54 on NSE and 56.04 on BSE, and HDFCBANK was 64.48 and 49.50.
So never mix exchanges in one series. BSE was 6.5% of combined turnover in this window. Its T series is 100% by construction, every one of 126,781 rows, but almost none of those pair with NSE EQ, so that does not explain the gap. BSE's value-weighted figure has also climbed from 45.24% in early 2025 to 64.08% so far this quarter while NSE stayed flat, and we have not explained that either.
The test
The rules:
- Universe: point-in-time top 500 by trailing turnover, May 2010 to November 2026, so delisted names stay in.
- Spike: turnover at least 3x the median of the prior 20 sessions. Turnover, because a split distorts share volume.
- Delivery: that day's % minus the stock's own prior 60-session mean, split into terciles within each year. Those cut-offs use the whole year, so they look ahead. We check that below.
- Return: adjusted, from the next day's open, since the file arrives after the close. Excess over the equal-weight universe for the same window.
- Noise: t-stats over monthly means, 195 months. Before all costs.
First surprise: spikes usually come with less delivery. Median delivery on a spike is 37.7%, against 45.5% across universe rows, and 7.18 points below the stock's own norm.
events median 5-day excess hit 20-day excess hit
deliv% (t) (t)
high 58,003 71.32 +0.114 ( 3.23) 47.2% +0.327 ( 3.41) 47.9%
middle 57,993 35.18 -0.455 (-9.71) 40.9% -0.509 (-5.26) 42.6%
low 57,989 24.87 -0.758 (-14.85) 39.0% -0.909 (-8.00) 41.9%
high - low +0.873 (14.68) +1.237 ( 8.52)We expected noise. This is not noise, and we tried to break it. The 5-day high-minus-low gap is +0.779 at a 2x spike and +0.933 at 5x, +0.676 entering at the next close instead of the next open, +0.874 without March 2020, and positive in 17 of 17 years. Setting the terciles from the prior year's cut-offs, so nothing looks ahead, gives +0.845 (t = 13.06) from 2011, against +0.874 on the same events. On days without a spike it is +0.128, so it is mostly a spike effect.
Why we still would not call it an edge
The long side is thin. High delivery beats the universe by 0.114% over five days before costs, and it beats it less than half the time. At a 5x spike it is -0.014 (t = -0.24). In 2023 to 2026 it is -0.073 (t = -1.10) at five days and -0.018 (t = -0.14) at twenty.
The gap is mostly the low side, and that is hard to reach. Acting on it means being short. Only 37,334 of the 173,985 events had a stock future that day. In those names the 5-day gap is +0.245 (t = 2.40), falling to +0.091 (t = 0.89) from the next close. In 2023 to 2026 it is +0.112 (t = 0.80), and -0.096 (t = -0.38) at twenty days. In the top 100 by turnover it is +0.140 (t = 0.79). The big number, +1.023, is in names with no future, at a median turnover rank of 353.
It has shrunk. The 5-day gap was +1.252 in 2010 to 2014, +0.851 in 2015 to 2019, +1.005 in 2020 to 2022 and +0.361 in 2023 to 2026.
The popular screen trails the universe. Up at least 5% on a 3x spike with delivery 10 points above the stock's norm: 1,419 events, -0.496 (t = -2.09) over five days and -0.775 (t = -1.77) over twenty, positive 42.0% of the time. With delivery of 60% or more instead: 1,899 events and -0.444 (t = -2.10). The same up-move with delivery 10 points below norm is -0.876 across 23,083 events. High delivery was less bad, not good. On this evidence, low delivery on a spike is a mild warning, and high delivery is not a buy signal.
We also ran a plain monthly sort on trailing 20-day delivery, top fifth against bottom fifth. It gives +1.016% a month (t = 3.15), but +0.725 (t = 1.32) in the top 100 and +0.396 (t = 0.61) in 2023 to 2026, and the top fifth is less liquid than the bottom fifth.
What this does not show: who was buying, why any of this happens, or that it survives costs, impact and borrowing, none of which we modelled. Excess is against an equal-weight benchmark. We ran more variants than are listed here (direction of the move, size of the move, absolute delivery level). They point the same way, but a t-stat from one of many tests means less than it looks. The 2,424 events with no 20-day exit, whether recent, delisted or suspended, are dropped, and they lean to the low side (967 low, 615 high). None of this is investment advice.
Reproduce it
Delivery, adjusted prices and the universe are all free on HuggingFace. This is the whole 5-day test, and it joins on (date, symbol):
import polars as pl
BASE = "hf://datasets/tejhq/indian-markets"
YEARS = range(2010, 2027)
adj = pl.concat([pl.read_parquet(f"{BASE}/prices_adjusted/nse_{y}.parquet", columns=[
"date", "symbol", "series", "open", "turnover", "adj_factor_cumulative", "adj_close"]) for y in YEARS])
dlv = pl.concat([pl.read_parquet(f"{BASE}/delivery/nse/year={y}/nse_{y}.parquet",
columns=["date", "symbol", "delivery_pct"]) for y in YEARS])
px = (adj.filter(pl.col("series") == "EQ").join(dlv, on=["date", "symbol"], how="left")
.with_columns(entry=pl.col("open") * pl.col("adj_factor_cumulative")))
sess = px.select(pl.col("date").unique().sort()).with_row_index("i")
px = px.join(sess, on="date").sort("symbol", "date").with_columns(
med20=pl.col("turnover").shift(1).rolling_median(20, min_samples=20).over("symbol"),
norm=pl.col("delivery_pct").shift(1).rolling_mean(60, min_samples=40).over("symbol"),
ret=pl.col("adj_close").pct_change().over("symbol"))
uni = pl.read_parquet(f"{BASE}/universe/nse_liquid.parquet").select("rebalance_date", "valid_to", "symbol")
month = sess.join_asof(uni.select("rebalance_date", "valid_to").unique().sort("rebalance_date"),
left_on="date", right_on="rebalance_date").filter(pl.col("date") <= pl.col("valid_to"))
m = (px.join(month.select("date", "rebalance_date"), on="date")
.join(uni.drop("valid_to"), on=["rebalance_date", "symbol"])
.join(px.select("symbol", pl.col("i") - 1, pl.col("entry").alias("buy")), on=["symbol", "i"], how="left")
.join(px.select("symbol", pl.col("i") - 5, pl.col("adj_close").alias("sell")), on=["symbol", "i"], how="left")
.with_columns(r5=pl.col("sell") / pl.col("buy") - 1)
.with_columns(xs=pl.col("r5") - pl.col("r5").mean().over("date"), year=pl.col("date").dt.year()))
ev = (m.filter(pl.col("turnover") >= 3 * pl.col("med20"), pl.col("ret").is_not_null(),
pl.col("delivery_pct").is_not_null(), pl.col("norm").is_not_null())
.with_columns(pct=((pl.col("delivery_pct") - pl.col("norm")).rank() / pl.len()).over("year")))
for name, cond in [("low", pl.col("pct") <= 1/3), ("high", pl.col("pct") > 2/3)]:
g = ev.filter(cond).group_by(pl.col("date").dt.strftime("%Y-%m")).agg(pl.col("xs").mean())
print(name, ev.filter(cond).height, round(g["xs"].mean() * 100, 3))
# low 58012 -0.759, high 58024 0.114 on data to 5 Nov 2026Or one symbol at a time through the API, with a free key:
from tej import Client
c = Client(api_key="tej_live_...")
nse = c.delivery("RELIANCE", exchange="nse", from_="2020-03-16", to="2020-03-25")
bse = c.delivery("RELIANCE", exchange="bse", from_="2026-10-01", to="2026-11-04")Dates outside coverage return no rows. If you do this differently and get a different answer, we would like to see it at github.com/tejhq/tej-bazaar. Rows join to nse/ or bse/on (date, symbol, series). For open interest on the same names, F&O is at /fo.