Reference run

Capacity & crowding on four reference factors

Synthetic NIFTY-100-like panel, five years, 100 names. Same engine you'd run on your own data via /analyze, plus the cross-sectional crowding signals that need universe-wide data.

Why this matters

Two compounding mechanisms degrade factor returns

Every factor strategy — momentum, value, quality, low-volatility — earns positive expected returns at small AUM and earns less at larger AUM. The usual story is "alpha decays as size grows." That story is incomplete. Two distinct mechanisms compound on each other:

Trading-cost drag. Costs scale super-linearly with size. Spread is roughly constant per unit of trade, but market impact grows like (Almgren–Chriss) and faster than that for large notionals. At some AUM the marginal alpha is fully consumed by impact.

Crowding. Other managers run the same signal. As capital piles in, the long basket bids up and the short basket gets squeezed. Realised alpha decays before trading costs even fire. Drawdowns become correlated across funds and unwinds become reflexive.

The PM-level question is not "what's my Sharpe at infinite AUM" — it is: given my current AUM and my factors' current crowding, what's my hard cap and what's my safe rebalancing rule?This monitor answers exactly that.

How it works

What the dashboard is computing in the background

Five stages run in sequence whenever the data is refreshed. Each stage's output drives one section of the dashboard below.

step 1

Build factor scores

Cross-sectional rank per factor: 12-1 momentum, P/B value, ROE quality, 6-month low-vol.

step 2

Construct portfolio

Top quintile long, bottom quintile short, dollar-neutral, equal-weighted.

step 3

Backtest gross

Daily gross-of-cost return series. Used as the alpha side of every capacity calculation.

step 4

Cost & capacity scan

At each AUM in the grid: spread + Almgren-Chriss impact + commission + borrow + India tax stack → Net IR.

step 5

Crowding, ML, policy

Composite crowding score, OU-model alpha forecast, regime classifier (GMM), optimal rebalance + buffer search.

Per-rebalance equations

is the bid-ask spread (Corwin-Schultz), trade dollars, annualised vol, daily ADV, chosen execution days, the impact coefficient, target weight, holding period.

ML layer · Gaussian mixture model

Current market regime

A 3-component GMM clusters daily market state (vol, average pairwise correlation, spread, ADV) into 'calm', 'normal', 'stressed'. The label drives whether nominal or stressed capacity should bind.

Today's regime

calm

argmax over the three GMM components

calm probability

93%

normal probability

stressed probability

Regime centroids (interpretable axes)

regime	market vol (ann.)	avg pairwise corr	median spread	log₁₀ ADV ($)
calm	0.235	0.319	1 bps	9.12
normal	0.251	0.353	1 bps	9.11
stressed	0.262	0.367	1 bps	9.17

Per-factor

Factor deep dive

Select a factor to inspect its capacity curve, cost decomposition, crowding score, and ML forecast.

Gross IR

+1.10

annualised, before costs

Safe capacity

—

Net IR ≥ 0.5 + zone constraints

Stress capacity

—

under tail-quantile market

Crowding

57amber

composite, 0-100, rolling-percentile

Best policy

Q / 0bps

net IR 0.58 at 10.0 cr

Capacity curve

Each point is the strategy run at a different AUM with trades scaled linearly. Y-axis is the Information Ratio after every cost; X-axis is AUM (log scale). The dashed line is the IR=0.5 PM hurdle.

Cost decomposition

Stacked annualised cost in bps, split by component. Spread, commission and India tax are roughly linear in turnover. Impact bends — square-root in participation up to the 10% of ADV threshold, then power-0.6.

Crowding score

Six external + internal signals each rolling-percentile-ranked into 0–100, then weighted into a composite. Higher = more crowded.

Latest snapshot

Component scores

signal	score	interpretation
valuation_spread	81	narrow long-vs-short P/B spread → factor mispricing arbitraged out (Asness)
alpha_decay	0	negative slope of trailing 1y IR → edge being competed away (Arnott)
short_interest	100	weighted SI on the short leg → squeeze risk and high borrow (Drechsler)
comomentum	2	long-leg names moving in lockstep → same arbitrageurs (Lou-Polk)
holdings_overlap	98	fraction of long names also held long by other factors (Sias)
internal_footprint	70	strategy's own liquidity strain at base AUM

Alpha-decay forecast

Six-month forward Sharpe projected via a mean-reverting state-space model fit on rolling 1-year IR. Crowding is the regime covariate — heavier crowding pulls the long-run attractor down.

half-life

512 days

deviation from attractor halves in this many days

crowding loading β

0.292

slope of attractor on crowding score

in-sample RMSE

0.099

daily increment residual, IR units

Operating policy search

Search over rebalance frequency × no-trade buffer at the safe-AUM target. The optimal point is highlighted.

freq	buffer (bps)	gross IR	net IR	cost (bps)	max exec days
D	0	1.12	-4.77	4651	3.0
D	25	1.12	-4.77	4651	3.0
D	50	1.12	-4.77	4651	3.0
D	100	1.12	-4.77	4651	3.0
D	200	1.12	-4.77	4651	3.0
W	0	1.14	-1.03	1714	2.5
W	25	1.14	-1.03	1714	2.5
W	50	1.14	-1.03	1714	2.5
W	100	1.14	-1.03	1714	2.5
W	200	1.14	-1.03	1714	2.5
BW	0	1.09	-0.38	1170	1.6
BW	25	1.09	-0.38	1170	1.6
BW	50	1.09	-0.38	1170	1.6
BW	100	1.09	-0.38	1170	1.6
BW	200	1.09	-0.38	1170	1.6
M	0	1.10	0.19	735	1.6
M	25	1.10	0.19	735	1.6
M	50	1.10	0.19	735	1.6
M	100	1.10	0.19	735	1.6
M	200	1.10	0.19	735	1.6
Q	0	1.12	0.58	435	1.3
Q	25	1.12	0.58	435	1.3
Q	50	1.12	0.58	435	1.3
Q	100	1.12	0.58	435	1.3
Q	200	1.12	0.58	435	1.3

Capacity checkpoints

As AUM utilisation climbs through 50% / 75% / 90% of safe capacity, execute these operational changes.

50% checkpoint

0.0 cr

Routine — keep current rebalance + buffer.

75% checkpoint

0.0 cr

Raise no-trade buffer to 100bps, downshift rebalance by one tier.

90% checkpoint

0.0 cr

Soft-close to new flows. Raise buffer to 200bps, switch to monthly rebalance.

Joint capacity

Multi-factor allocation

Total target AUM split across factors to maximise expected net return, constrained by the crowding red line.

Total AUM

5.0 cr

2× max single-factor safe capacity

Expected net return

11.9%

annualised, after all costs

Selected factors

1/4

non-zero allocations

factor	weight	AUM	net IR	crowding
momentum	0.0%	0.0 cr	0.73	57
value	100.0%	5.0 cr	0.95	38
quality	0.0%	0.0 cr	0.64	44
low_vol	0.0%	0.0 cr	-0.14	48

ML layer · ridge regression

Capacity prediction model

A small ridge regressor that lets you predict 'safe AUM' from market state without re-running the full curve. Lambda picked by leave-one-out CV.

In-sample R²

0.01

how much variance the linear model explains

Selected λ

100.00

ridge regularisation, picked by LOO-CV

Features

spread, vol, ADV, crowding, gross IR

feature	β coefficient	interpretation
spread_bps	0.0000	+1 bp of spread increases log-AUM by 0.000
vol	0.0000	+1pt vol increases log-AUM by 0.000
log_adv	-0.0004	+1 in log10(ADV) decreases log-AUM by -0.000
crowding	0.0428	+1pt crowding score increases log-AUM by 0.043
gross_ir	0.0972	+1 unit of gross IR increases log-AUM by 0.097
intercept	1.4497	log₁₀ AUM at zero feature values