funding-rate-alpha

KO | EN

gitlite — search

#python #react #javascript #machine-learning #golang #pytorch #python3 #flutter #c #component-library #css #deep-learning

funding-rate-alpha

★ 39 Open GitHub ↗

Do funding rates predict returns? A survivorship-free test on 6.1M settlements across three venues.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Funding Rates as a Cross-Sectional Factor in Perpetual Futures

Structural processing, multi-dimensional cross-validation, and the
tradeable residue of funding-rate signals on the complete USDT-margined
perpetual record of a major venue, 2020 through mid-2026.

---

## Abstract

A public practitioner note proposed that funding-rate trading factors
require per-contract rolling z-score standardization, cross-sectional
rank ordering, cross-validation with open-interest change and with
price, and swing rather than daily rebalancing. Each claim is tested on
a venue's complete USDT-perpetual record: 791 contracts including every
delisted one, 2.43 million funding settlements, hourly prices, and
5-minute open-interest metrics for a survivorship-free union of 421
liquid contracts, extended to 6.1 million settlements across three
venues. The funding formula is verified against the premium index
archives at 0.61 basis points median error.

The raw funding level is a carry trade far more than a price signal. The
dollar-neutral decile book collects 11.6 bps per day of funding, two
thirds of it from the long leg rather than from shorting crowded
contracts, while the price coefficient changes sign between a 2020-21
reversal regime and a 2023-26 continuation regime. Tested across ten
horizons with Newey-West errors, funding does predict returns, in the
level and not the z-score, at eight hours and shorter. That pooled
result averages two regimes: nothing before 2024 except a significantly
negative coefficient at a day, and a large effect at every horizon
after, growing as the interest-rate peg stopped binding. Open interest
builds into and through high-funding settlements, which refutes a flow
reading and leaves the rate as a measurement of the basis.

No cross-validation prescription survives a control. The open-interest
conditioning collapses from a pooled $+34$ against $+6$ bps per day to
$+7.0$ at $t = 0.35$ once cohorts are compared inside the same day;
open-interest change is itself largely short-horizon reversal measured a
second way; no composite adds anything to the funding z-score it
contains; and the open-interest-weighted
cross-venue index returns the dominant venue's own rate at a correlation
of 0.975. What remains is carry harvesting, netting 8.9 bps per day
over the full sample at a 4 bps one-way schedule, but earning two thirds
of that in 2020-21 and not clearing a threshold thereafter. It breaks
even only at a fee of 23 bps, so cost is not what limits it. Its capacity is set by its smallest
holding at roughly USD 17M per leg, and raising the liquidity floor
removes the price leg rather than scaling it. Neither carry-to-risk
ranking, inverse-volatility sizing nor factor hedging recovers any of
it, the book's beta to the common perpetual factor being $+0.013$, and
holding only across the settlement timestamp fails on 1.73 million
settlements because the price moves against the collector by more than
the payment.
## 1. Introduction

Perpetual futures pay a periodic funding rate that transfers value
between longs and shorts and anchors the contract price to an index.
Because retail-favorite contracts carry structurally high funding, the
cross-sectional level of funding measures how retail-oriented a contract
is rather than how crowded its current position structure is. The source
note (a public post, status 2085417405875167278, archived in the shared
data store; English rendering in `docs/source_tweet.md`) prescribes
rolling per-contract z-scores before cross-sectional ranking, open-
interest and price cross-validation, and swing-horizon execution.

Each prescription is tested on full history including delisted
contracts, with time-varying tradability filters, taker and maker cost
accounting, tail and capacity analysis, and placebo controls. Funding
mechanics and their verification are in `docs/funding_math.md`, the
cross-venue archives in `docs/venues.md`.

## 2. Data and method

Table 1 lists the archives the study is built on.

**Table 1.** Public archives used, with coverage and contract counts.

| Set | Coverage | Source |
|---|---|---|
| Funding settlements | 2020-01 to 2026-06, 791 contracts, 2.43M events | monthly fundingRate archives |
| Hourly klines | 2020-01 to 2026-07, 788 contracts | monthly/daily kline archives |
| Open interest and position ratios | 2020-09 to 2026-07, 421-contract download universe, 438 parsed, 420 in the panel | daily metrics archives, 5-min snapshots aggregated to hourly |
| Premium index klines | 2020-01 to 2026-06, top 50, 1h and 8h | monthly premiumIndexKlines archives |

All archives are public and frozen by manifest in `manifests/`. Parsed
datasets are stored on the shared data volume, whose location is read from
the `FUNDING_DATA_ROOT` environment variable (see `scripts/_paths.py`):
`funding_events.parquet`, `metrics_hourly.parquet` and
`decision_panel.parquet`, a daily 00:00 UTC panel of signals, open-interest
features, price features and forward outcomes. The funding archive holds 791
contracts; the 671 in the analysis files are those with hourly
prices, a known forward return and at least one day inside the tradability
filter, which requires a trailing 90-day mean daily quote volume of at least
USD 5M evaluated at each decision time; open-interest analyses further
restrict to the top 150 by trailing 365-day volume, evaluated causally.

Table 2 maps the source note's claims onto the sections that test them.

**Table 2.** The six hypotheses and where each is answered.

| | Claim | Tested in |
|---|---|---|
| H1 | Raw cross-sectional funding measures retail orientation, not crowding | 4 |
| H2 | Rolling per-contract z-scores, then cross-sectional rank, improve it | 5 |
| H3 | High funding with rising open interest marks accumulating crowding | 8 |
| H4 | High funding with flat or falling price marks stuck longs | 8 |
| H5 | A cross-venue open-interest-weighted index, and dispersion as sentiment | 10 |
| H6 | Fast decay favours swing over daily rebalancing | 6, 12 |

Open-interest coverage is uneven and worth stating. The metrics
archives begin in September 2020 and are backfilled per contract, so
only 2% of tradable symbol-days in 2020 and 11% in 2021 carry open
interest, against 99% in 2022 and 72% to 85% thereafter as the universe
outgrew the download set. Over the whole sample it is 75%, but every
open-interest analysis runs on the causal top-150, where coverage is
91%. Missing rows are dropped rather than filled, so those sections are
measured on the days the data exists.

Returns and execution are constructed as follows. Decision times are
00:00 UTC daily. The price at an hour boundary is the close of the
candle opening one hour earlier, so a forward return over $h$ hours is
the close $h$ candles later divided by the close at the boundary, minus
one. Funding carry over a holding window is the signed sum of the rates
of every settlement falling strictly inside it, positive to a short.
Rebalances are priced at the decision-time close, and cost is charged as
a one-way fee on the churned fraction of each leg, doubled for the round
trip. Net figures use a reference schedule of 4 bps one-way taker unless
a maker schedule of 2 bps is named. That is a reference point rather than
a premise the conclusions rest on: published futures fees vary by venue
and by volume tier, so Section 12 reports each strategy against the whole
fee curve and gives the fee at which it breaks even.

Three conventions govern every statistic below.

**Minimum cross-section.** Daily information coefficients, quantile sorts
and portfolio returns are formed only on days carrying at least 20 tradable
names, except in two Section 7 subsamples too narrow to support it, which
use eight and are flagged where they appear. The cross-section runs from 3
to 381 names, median 153, and a 3-name day yields a rank correlation of $\pm
1$ or $\pm 0.5$ that would otherwise weigh as much as a 381-name day. The
rule drops 77 of 2281 days, all in 2020.

**Holding-period marking.** Multi-day baskets are chosen inside the tradable
universe but marked against the unfiltered panel for the whole holding
period, with a fixed denominator equal to the basket size. A name that falls
below the liquidity filter or delists mid-hold therefore realizes its move
instead of being dropped and its weight redistributed to the survivors.
Coverage of basket-days is 99.94%.

**Cost booking.** The round-trip cost of a rebalance is charged once, in
full, on the first day the new basket is held.

## 3. The funding mechanism

The venue charges, at each settlement timestamp,

$$F = \bar{P} + \mathrm{clip}(r_I - \bar{P},\, -c,\, +c), \quad
r_I = 0.0001, \quad c = 0.0005,$$

where $\bar{P}$ is the interval average of the per-second premium index,
the median of order-book imbalance prices and the last trade against a
spot basket index. When $|\bar{P} - r_I| \le c$ the rate is pinned at the
interest rate, $+0.01\%$ per 8h, about $+10.95\%$ annualized paid by
longs: the structural baseline the z-score must remove. Settlements
occur every 8h by default; 541 of the 791 contracts run on a sub-8h
schedule at some point: 394 are listed that way from the start and 147
move off the 8h schedule during their life. The
panel works from the actual event stream and every annualization uses
each event's own interval.

Figure 1 verifies the formula: applying it to the interval-averaged
premium reproduces the archived rates at a median per-contract mean
absolute error of 0.61 bps (0.24 bps on the largest contract), against
2.5x worse for the terminal-sample premium, which identifies the
averaging convention. Caps bind well above the default: the largest
observed rate is 4.00% per interval, 42 contracts exceed 2% and 408 of
791 exceed 0.75%, so the venue runs several cap tiers rather than the
one the documentation describes.

![Figure 1](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig1_formula_verification.png)

**Figure 1.** Funding formula verification: archived settlement rates
against the formula applied to the interval-averaged premium index for
three contracts, with the 45-degree line; the legend reports each
contract's mean absolute error in basis points, the statistic tabulated
per contract in `results/premium_check.txt`.

## 4. Baseline effect

Figure 2 shows the funding landscape. The cross-sectional median sits on
the interest-rate peg in quiet years and departs sharply in 2021 and
2024; Figure 2 draws the interquartile band, and pooled across all
contract-days the 10th-to-90th percentile range spans $-17\%$ to
$+17\%$ annualized. Concentration at the peg is the dominant feature and it
decays through the sample: the share of contract-days priced exactly at
the interest rate runs 58%, 52%, 47%, 61%, 41%, 21%, 7% from 2020 to
2026, and the cross-sectional interquartile range narrows to about 1%
annualized in 2023 and to zero in 2024.

![Figure 2](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig2_funding_regimes.png)

**Figure 2.** Cross-sectional funding rate, all liquid USDT perpetuals:
median and interquartile band, annualized, with the interest-rate peg.

Sorting on the raw rate (Figure 3) separates carry from price. The
short-side carry is monotone in the rate, $-11.0$ to $+4.9$ bps per day
across quintiles, while next-24h price returns are non-monotone noise.
A daily long-short decile on the raw rate nets $+17.7$ bps per day
(Sharpe 2.40 after 4 bps taker one-way, clipped forwards), decomposing
into a carry leg of $+11.6$ bps per day and a price leg of $+10.2$ bps
per day whose drawdown reaches $-74\%$. The carry is not where the
crowding story puts it. At full notional the short leg collects $+7.5$
bps per day and the long leg $+15.7$, so 68% of the book's carry comes
from being long contracts paying deeply negative funding, where shorts
pay longs, rather than from shorting the retail-crowded side. Figure 3b
shows the asymmetry directly: the bottom quintile pays a short $-11.0$
bps per day while the top pays only $+4.9$. The carry leg accrues
almost without variance, a ratio of return to volatility above 17, but
that is a property of an accrual rather than a tradeable Sharpe: it is
collectable only by holding the price leg that carries the risk. By year
the net
return is $+228$, $+216$, $+50$, $+16$, $-14$, $+20$, $-43$ percent
(2020 to 2026). The raw level is therefore a carry harvest with
regime-dependent price risk, not a price signal: its 24h information
coefficient is $+0.0023$ ($t = 0.9$) while the 8h coefficient is
$+0.0104$ ($t = 4.0$), short-horizon continuation only.

![Figure 3](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig3_baseline_quintiles.png)

**Figure 3.** Baseline effect of sorting on the raw funding rate, with
per-name forward returns clipped to $[-90\%, +200\%]$ as in the second
quintile table of `results/baseline_carry.txt`. (a) Next-24h price
return by quintile, annualized. (b) Carry collected by shorts and total
short-side return, bps per day.

## 5. The regime flip

Figure 4 is the central fact of the study. Rolling 252-day Spearman
coefficients of the funding level against next-24h returns are negative
in 2020-21 (high funding preceded reversals) and positive from 2023
(high funding preceded continuation). Table 3 gives the yearly
coefficients for both the level and its z-score; the level is the
stronger of the two in every year from 2024, which Section 6 takes up.

**Table 3.** Yearly Spearman information coefficient with t-statistics:
the raw funding level and its 90-day z-score against the next-24h return,
and the raw level against the next-1h return, the horizon at which it is
strongest. The 24h rows carry the naive t-statistic used elsewhere in
this report; the 1h row carries the Newey-West correction of Section 6.
The correction is not uniformly the stricter of the two -- it shrinks the
statistic where the daily coefficient is positively autocorrelated and
raises it where it is not -- so both are tabulated in
`results/zscore_factor.txt`. 2020 and 2026 are partial years.

| year | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|---|---|
| raw level 24h, IC (%) | -3.9 | -3.1 | +0.0 | +0.6 | +1.6 | +3.4 | +2.0 |
| raw level 24h, t | -2.5 | -4.7 | +0.1 | +1.1 | +3.4 | +8.1 | +3.6 |
| z-score 24h, IC (%) | -2.3 | -3.2 | -0.3 | +0.8 | +1.1 | +2.1 | +0.5 |
| z-score 24h, t | -1.5 | -5.1 | -0.4 | +1.4 | +2.4 | +4.5 | +0.8 |
| raw level 1h, IC (%) | -1.1 | +1.0 | +2.0 | -1.3 | +2.2 | +5.2 | +5.9 |
| raw level 1h, t | -0.7 | +1.7 | +2.6 | -1.9 | +3.8 | +12.7 | +7.2 |

![Figure 4](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig4_regime_flip_ic.png)

**Figure 4.** Rolling 252-day Spearman information coefficient for the
raw rate and the 90-day z-score, on shared axes. (a) Against the next-24h
return, with the 2020-21 reversal regime and the 2023 to mid-2025
continuation regime shaded. (b) Against the next-1h return, the horizon
at which the level is strongest. The shading marks 2024 onward, which is
where the raw level's yearly coefficient stops alternating in sign rather
than where the effect first appears; the z-score plotted beside it still
crosses zero inside that band.

Panel (b) repeats the measurement at one hour, which Section 6
identifies as the horizon where the level is strongest. The pairing is
the point: the flip in panel (a) happens at a horizon whose pooled
coefficient is not distinguishable from zero ($+0.23\%$, $t = 0.79$),
while the one-hour coefficient is ($+1.89\%$, $t = 6.26$). Nor is the
one-hour effect a clean switch. Its yearly coefficient alternates sign
through 2023, so it pools to nothing over 2020-23, and is then positive
in every year from 2024, rising from $+2.2$ to $+5.9$. The doubling
twice belongs to the decile spread rather than the coefficient, and is
taken up in Section 6.

No unconditional level signal survives this flip. The z-score
($z_{s,t} = (r_{s,t} - \mu_{s,W}) / \sigma_{s,W}$ over a rolling 90-day
window of the contract's own settlements, then cross-sectional rank)
stabilizes the strategy across regimes: the daily z-score decile is
positive in six of seven years against five of seven for the raw level,
and its price leg draws down $-50\%$ against the raw level's $-74\%$. The
24h IC of the z-score itself remains indistinguishable from zero
($-0.0007$, $t = -0.3$): the preprocessing removes the baseline without
creating predictability.

## 6. What funding predicts

Sections 4 and 5 have shown what the funding level does at a day: it
pays carry, and its price coefficient changes sign between regimes. That
leaves the question the report exists to answer still open, because a
single horizon cannot answer it. This section tests ten, from one hour
to thirty days, with Newey-West standard errors, since a daily
coefficient series is autocorrelated and overlapping windows make the
ordinary statistic too generous.

Pooled over the sample, funding predicts returns only in the level, only at
eight hours and shorter, and only by too little to trade. The raw rate
scores $+1.89\%$ at one hour ($t = 6.26$), decaying to $+1.04\%$ at eight
(Figure 5a), and is gone by twenty-four ($+0.23\%$, $t = 0.79$). Thirty
shuffled-signal placebos at one hour return $-0.020\%$ with one of thirty
exceeding $|t| = 2$, and thirty at twenty-four hours return $+0.004\%$ with
two of thirty, both the rate chance predicts. The 90-day z-score predicts
nothing at any horizon, $+0.00\%$ at one hour. That sharpens Section 5
rather than repeating it: the preprocessing does not merely fail to create
predictability, it removes the only component that has any.

Two qualifications matter more than the pooled result.

The four short horizons are one finding, not four. They are nested
windows from a common origin, so removing the first hour from each is
the test of whether anything happens after it. Pooled, nothing does:
$-0.03\%$ at two hours, $-0.01\%$ at four, $+0.47\%$ ($t = 1.60$) at
eight. The apparent monotone decay from 1.89 to 1.04 is one hour of signal
diluted across a widening window, so the four Benjamini-Hochberg
survivors are one discovery restated, not four.

And the whole result belongs to the second half of the sample. The regime
flip of Section 5 governs the short horizon as much as the daily one.
Split at 2024:

| | 1h | 2h | 4h | 8h | 24h |
|---|---|---|---|---|---|
| 2020-2023 | +0.33 | -0.30 | -0.30 | -0.29 | **-1.29** |
| 2024-2026 | **+4.10** | **+3.61** | **+3.05** | **+2.93** | **+2.39** |

Over the first four years, 59% of the sample, there is no effect at any
horizon except a significantly negative one at a day ($t = -3.0$). From
2024 the effect is large and significant everywhere, including the
twenty-four hours where the pooled figure reads zero ($t = 8.8$). The
pooled $+0.23\%$ is the average of two opposite significant regimes and
means nothing on its own. In the modern era it is also
no longer confined to the first hour: with that hour removed, two-hour
through twenty-four-hour coefficients run $+0.97\%$ to $+1.40\%$ at
$t = 2.9$ to $5.1$.

Pricing it must be split the same way, and the direction has to be stated
because it changes with the horizon. At one hour the coefficient is
positive, so the tradeable direction is long high funding, and the decile
spread is 3.66 bps in that direction against the 8 bps reference round
trip. At
one to three days the spread in Figure 5b turns positive in the opposite
convention, reaching $+10.2$ bps at a day and higher at two and three, which
looks like a trade that clears its cost. It is not this section's effect:
that is the carry and reversal price leg of Section 4, which points the
other way, is concentrated in 2020-21, and whose excess over the round trip
is itself indistinguishable from zero ($+2.2$ bps, $t = 0.68$ at a day). The two
horizons carry opposite signs and only the short one is the subject here. By
year the one-hour spread runs $+0.9$, $+0.5$, $-1.1$, $+0.7$, $+4.5$, $+8.8$, $+16.7$ bps from
2020 to 2026. For 2025-26 it is $+11.4$ bps ($t = 8.8$) per round trip, enough to cover
a one-way fee of 5.7 bps; over 2024-26 the spread is $+8.6$ bps ($t = 9.3$)
and covers 4.3 bps. Over 2020-23 it covers 0.07 bps, which is to say
nothing. Whether the effect pays for itself therefore turns on the
schedule and not on the sample: it is absent in the early years and, in
the recent window, clears the fee at any competitive tier. The growth
tracks the collapse of the interest-rate peg
documented in Section 4: the share of contract-days pinned exactly at it
falls from 61% in 2023 to 7% in 2026, and the rate only carries
cross-sectional information once it stops being a constant. That estimate
assumes execution at the settlement print itself and is an upper bound; how
quickly it decays intraday is not something hourly bars can answer.

One thing genuinely is predictable: the regime. Regressing each day's
coefficient on its own trailing average, with a further day of embargo
so that no trailing observation overlaps the outcome, gives a slope of
$+0.65$ to $+0.78$ across 63, 126 and 252-day windows at $t = 5.7$ to
$6.6$, and the trailing sign matches the next day's 54.7% to 56.3% of
the time. Knowing the regime is not the same as profiting from it: a
strategy trading the sign of its own trailing coefficient earns $+1.9$ to
$+2.8$ bps per day more than always shorting high funding, but at
$t = 0.64$ to $0.88$ and a block-bootstrap $p$ of 0.19 to 0.29 that gain
is not distinguishable from zero.

![Figure 5](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig5_predictability.png)

**Figure 5.** Predictability by horizon, pooled over the sample.
(a) Spearman information coefficient of the raw rate and the 90-day
z-score against forward returns from one hour to thirty days, with 95%
Newey-West bands. (b) The decile spread from shorting high funding at
each holding period against the 8 bps taker round trip. Negative bars
are the one-hour effect of this section, whose tradeable direction is
the opposite one; the positive bars at one to three days are the carry
and reversal price leg of Section 4, a different effect pointing the
other way. Both panels pool 2020-2026 and so average the two regimes
tabulated above.

## 7. What the one-hour effect is

The one-hour effect of the previous section is a fact without a
mechanism, which is a fact waiting to be explained away. Four candidates
are separable here, because each predicts something the others do
not.

Two fall immediately. It is not beta: the premise usually
offered, that high-funding contracts are the volatile retail-heavy
names, is not true on this panel, where the within-day rank correlation
of the rate with 7-day volatility is $-0.013$. It is not stale
momentum: on the cross-section at 00:00 UTC the funding slope moves from
$+0.0201$ alone to $+0.0150$ ($t = 4.37$) with trailing one-hour and
eight-hour returns alongside, an attenuation of a quarter rather than an
elimination, and the one-hour reversal it survives is powerful
($-0.080$, $t = -20.1$).

The settlement clock matters, but not cleanly. Rebuilding the measurement at
all 24 hour boundaries turns a constant into a sawtooth (Figure 6a): hours
that begin a settlement average $+0.75\%$ against $-0.04\%$ for the eighteen
that do not, a difference of $+0.79\%$ at Welch $t = 2.62$. Correcting the
24 hourly tests for multiplicity leaves six survivors, of which only three
are settlement hours; 06:00, 17:00 and 23:00 also survive, two of them
negative, and excluding 00:00 the eight-hour-grid contrast collapses. The
clock is visible in the data and 00:00 UTC carries most of it, but the
effect is not confined to settlement hours.

A flow reading is refuted outright by the variable that separates it
from the alternative. If longs cut before the payment and rebuild after,
open interest must fall into a high-funding settlement and recover after
it. If instead shorts are opening to collect the payment, it rises into
the print. The price path is the same under both; open interest is not.
Across 685,816 settlements on 206 contracts, the rank correlation of the
rate with the change in open interest is $+1.73\%$ ($t = 9.90$) over the
hour ending at the settlement and $+1.12\%$ ($t = 7.07$) over the hour
beginning at it. Open interest builds into a high-funding settlement and
keeps building afterwards: there is no round trip, and the sign is the
opposite of what longs closing would produce. Positions are opened into
the payment, not unwound before it.

Nor is the effect about the imminent payment at all. A rate lagged a
full week predicts the next hour as well as the one that has just
settled: $+1.77\%$ ($t = 6.65$) against $+1.89\%$ ($t = 6.26$), and the
one-day lag is higher than either at $+2.09\%$. Whatever the rate is measuring
changes slowly, which is what a persistent property of a contract looks
like and not what an imminent cash flow looks like.

What survives is narrower and duller than a flow story. The rate is a
clipped average of the premium index, so it is a measurement of how rich a
perpetual trades against spot, and on the 36 contracts where both are
available the contemporaneous premium subsumes the rate that summarises it,
keeping $t = -2.07$ against the rate's $-1.49$ when both are carried. That subsample
carries a caveat large enough to state plainly: those 36 are the largest
contracts in the sample, and on them the one-hour effect runs the other way,
$-1.32\%$ ($t = -2.13$) against $+1.89\%$ on the full panel. The premium subsuming the rate is
evidence about which of the two is the more direct measurement, not
confirmation that the basis drives the effect at large. The basis reading is
where the evidence points; it is not established. The basis is persistent,
it carries short-horizon information, and the settlement is simply when the
published measurement of it refreshes. That reading is consistent with the
size gradient, the coefficient running $+2.38\%$ and $+3.59\%$ in the smallest and
middle thirds by volume against $+0.69\%$ ($t = 1.76$) in the largest (Figure 6b, on a
floor of eight names per cross-section rather than the usual twenty), and
with the effect appearing only once funding stopped being pinned at the peg.
It leaves two things unexplained: why 00:00 UTC should carry nearly twice
the next-largest settlement hour and more than three times their average,
and an even-hour against odd-hour gap of $+0.71\%$ ($t = 6.3$) that persists among
hours on no funding grid at all ($+0.47\%$, $t = 3.65$).

![Figure 6](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig6_mechanism.png)

**Figure 6.** What the one-hour effect is. (a) The one-hour coefficient
by hour of day with 95% Newey-West intervals, coloured by whether the
hour begins a settlement on the 8-hour grid, the 4-hour grid only, or
neither. (b) The same coefficient by trailing-volume tercile. Both pool
2020-2026; the effect is concentrated in 2024 onward.

## 8. Cross-validation with open interest and price

Sections 6 and 7 settle what the funding rate alone predicts. The source
note's remaining claims are that it predicts more in combination: with
open-interest change, with price, and across venues. Those are taken in
turn here and in the two sections that follow.

Open-interest change over 24h has a real univariate coefficient, IC
$-0.022$ ($t = -7.0$): building open interest precedes underperformance,
though Section 11 shows most of that is short-horizon reversal measured
a second way. Its prescribed role as a conditioner of funding is where
the source note fails.

The multiplicative interaction with the funding z-score is weak on its
own ($t = -0.9$). The quadrant comparison looks stronger (Figure 7b):
pooled over all observations, high-funding contracts return $+34.1$ bps
per day when open interest builds against $+5.9$ when it unwinds. Those
cohorts are drawn from different days: open interest builds market-wide
on some days and unwinds on others, so the pooled gap mixes the cohort
effect with a day effect. Differencing the same two
cohorts inside each day and averaging over the 531 days on which both
exist leaves $+7.0$ bps per day at $t = 0.35$, and the sign is unstable
by year ($+22$, $-18$, $-4$, $-7$, $+43$ for 2022 to 2026).

![Figure 7](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig7_oi_sorts.png)

**Figure 7.** Open-interest cross-validation. (a) Next-24h return in bps
per day by funding tercile and 24h open-interest-change tercile.
(b) High-funding quadrants (z-score at least one), pooled across days;
the annotation reports the same contrast computed within each day, which
is the statistic the conclusion rests on.

The price dimension (Figure 8) fails the same way. Plain 24h reversal is
the strongest single predictor in the study, IC $-0.041$ ($t = -10.8$).
The stuck-long state (z-score above 1 while price is flat or down) shows
pooled continuation of $+15$ bps per day, but against day-matched
size-controlled placebo cohorts the excess is not significant
($p = 0.23$, 200 re-draws per day), the within-day contrast against
high-funding-and-rising contracts is $-17.7$ bps per day at $t = -0.63$,
and the state is regime-split by year ($+92$ in 2020 to $-51$ in 2026).

![Figure 8](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig8_price_sorts.png)

**Figure 8.** Price cross-validation. (a) Next-24h return by funding
tercile and trailing 24h return tercile. (b) Stuck-long state
(z-score at least one with non-positive trailing return) by year.

## 9. Composites

The multiplicative composite prescribed by the source note, with weights fit
on the 393 days to the end of 2022 and tested 2023 onward, selects 0.75 on
the plain funding z-score, 0.25 on the price interaction and zero on the
open-interest interaction. Its train IC is $-0.001$ ($t = -0.1$) and its
test IC $+0.012$ ($t = 3.9$), but that out-of-sample coefficient is the
funding z-score's own, not a gain from the interactions: on the same test
window and universe the z-score alone scores $+0.011$ ($t = 3.6$), reported
in `results/composite.txt`. What the number measures is the continuation
regime of Section 5. The long-short top/bottom-20% portfolio nets $+2.2$ bps
per day at taker fees, Sharpe 0.59, against 57% daily churn.

An additive rank composite of open-interest change, reversal and funding
z-score reaches IC $-0.026$ ($t = -7.9$), and dropping the funding leg
improves it to $-0.031$ ($t = -8.7$). Two of those three legs turn out to be
one: Section 11 shows open-interest change is largely reversal measured a
second way, so the composite is less diversified than it looks. Neither
survives daily rebalancing: the three-signal version loses $-3.4$ bps per
day at taker fees, where churn of about 76% of each leg consumes the edge.

## 10. Cross-venue: the synthetic global index and dispersion

The third of the note's cross-validation claims is that a single venue's
rate is the wrong object: the rigorous construction is an
open-interest-weighted synthetic global funding index, and the cross-venue
dispersion of funding is itself a sentiment signal. Testing it required
funding archives from two further venues (`docs/venues.md`): 1.14M
settlements for 521 contracts from 2022 and 2.55M for 847 contracts from
2020, both including delisted contracts recovered by probing names rather
than reading a live instrument list, which adds 98 and 158 contracts
respectively. Signals may use every venue; the payoff is always Binance
price return plus Binance funding carry, the venue the position is held on.

Venues do disagree, as claimed. Pairwise correlations of the annualized
rate run 0.82 to 0.84, the median absolute gap between two venues on the
same contract-day is 8.6% annualized (0.0% for the pair that most often
sits pinned together at the peg), and the 90th percentile of that gap
reaches 33% to 38% annualized (Figure 9).

![Figure 9](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig9_venue_agreement.png)

**Figure 9.** Cross-venue agreement. (a) The annualized funding rate of
the largest contract on each venue, 14-day mean, with the interest-rate
peg. (b) Distribution of the absolute gap between venue pairs across all
contract-days, outliers suppressed.

It does not help. The equal-weighted global index posts a higher
headline than the single venue on a three-day swing, 6.0 against 4.4 bps
per day, but the two trade the same days and mostly the same names, so
the difference has to be tested paired. Differenced day by day, it
earns $+1.7$ bps per day at $t = 0.77$, and the entire
advantage is 2021, the year in which only 23% of contract-days carry a
second quote at all and the third venue has no archive whatsoever
(Figure 10b). From 2022, when all three venues are present, the sign
reverses to $-1.0$ bps per day at $t = -0.56$. Weighting by open
interest, the specific refinement the note prescribes, makes the index
worse rather than better on both schedules ($-1.4$ bps per day against
equal weight, $t = -0.53$).

![Figure 10](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig10_global_index.png)

**Figure 10.** The synthetic global index against the single venue.
(a) Three-day swing equity for the single-venue z-score, the
equal-weighted global index and the open-interest-weighted index, net of
4 bps taker one-way, log scale. (b) Cumulative day-by-day difference
between the global index and the single venue: the advantage is earned
in 2021, when multi-venue coverage is thinnest, and given back
afterwards.

Dispersion looks like a signal and is not one. Its raw cross-sectional
coefficient is $-0.012$ ($t = -4.6$), which would be the strongest
result in this section. But venues disagree most about contracts whose
funding is extreme in either direction, and dispersion ranks 0.23
against the absolute funding level. Removing the funding level within
each day leaves $-0.004$ ($t = -1.5$), and a double sort puts the whole
effect in one corner: high-minus-low dispersion is $-32$ bps per day
inside the low-funding tercile and $+1$ and $-1$ bps in the other two.
It is also untradeable in the direction the sentiment reading implies,
losing 10.0 bps per day at daily rebalance and 4.6 on a three-day swing,
because shorting the high-dispersion decile means paying funding rather
than collecting it. The time-series form fares no better: market-wide
disagreement carries a correlation of $+0.02$ ($t = 1.0$) with the next
day's return, and the high-minus-low-disagreement day difference is
$+27$ bps at $t = 0.68$.

The one relationship that survives its controls is a venue basis: funding
rich on one venue relative to the others precedes mild outperformance,
$+0.009$ ($t = 3.5$) for the venue the returns are measured on once its own
funding level is removed, and $+0.006$ ($t = 2.4$) and $-0.003$ ($t = -1.0$)
for the other two. At an information coefficient below 0.01 it is
economically negligible, and 78% churn per rebalance buries it: the
long-short loses 6.6 bps per day at daily rebalance and 2.5 on a three-day
swing.

## 11. Positioning level

A practitioner replying to the source note (`docs/source_tweet.md`)
raised three objections, each testable on this data and answered here
under the report's own conventions.

The first supplies the mechanism behind a result in Section 10.
Funding and positioning share a source, the argument runs, so cross-venue
differences are small residuals on one latent variable; and because one
venue dominates open interest, weighting a synthetic index by open
interest returns that venue's own rate. That is measurable. On the
273,762 base-days carrying both venues' open interest the dominant
venue's weight has a median of 66.5%, and the open-interest-weighted
index correlates with that venue's own rate at **0.975**. Section 10
found that weighting by open interest made the index worse rather
than better; this is why. The weighted index is the dominant venue plus
estimation noise, so it can only underperform the venue it is mostly
made of.

The second objection is that funding is a rent to be collected rather
than a signal to be optimised, and that cross-sectional alpha lives in
the positioning level instead. The relevant series is the long/short account
ratio of top traders, which covers 65% of tradable symbol-days across
409 contracts from September 2020. It does
predict: an information coefficient of $+0.013$ ($t = 4.7$), surviving
controls for open-interest change and trailing return at $+0.013$
($t = 4.4$). The sign is positive, so contracts where top traders are
net long outperform, which is continuation rather than the reversal a
crowding reading implies. Which signal wins depends on the criterion, and the
two disagree (Figure 11a). On the coefficient the funding level leads at
$+0.020$ ($t = 7.5$) against positioning's $+0.013$, with the funding
z-score last at $+0.011$; all three are weaker than trailing reversal.
Traded as decile long-shorts on the same days the order inverts
completely: positioning nets $+4.2$ bps per day at Sharpe 0.82, the
funding level $+2.8$ at 0.51 and the z-score $+1.4$ at 0.27. The reason
is turnover, 19% of each leg per day against 49% and 61%. Positioning is
a real signal whose advantage is that it moves slowly, not that it
predicts better.

The third objection is to keep the construction simple, with no
open-interest or price overlay and a weekly rebalance. The first half is
supported and sharpens Section 8. Carried together with trailing reversal on
the full panel, the open-interest slope falls from $-0.024$ ($t = -8.3$) to $-0.005$ ($t = -1.9$)
while reversal is untouched at $-0.036$ ($t = -8.5$). What Section 8 records as a
standalone crowding signal on its univariate coefficient is largely
short-horizon reversal under another name, which also makes the composites
of Section 9 less diversified than their three legs suggest. The weekly half
is not supported. Sweeping the holding period of the funding z-score across
the full sample gives net returns of $+10.4$, $+8.9$, $+9.4$, $+3.3$, $+5.3$ and $+5.7$ bps
per day at one, three, five, seven, ten and fourteen days, and Sharpe of
1.55, 1.42, 1.53, 0.57, 0.88 and 0.96. The sweep is flat from one to five
days and falls after; a weekly rebalance is the worst cell in it (Figure
11b).

![Figure 11](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig11_positioning.png)

**Figure 11.** The two objections that turn on a criterion.
(a) Cross-sectional coefficient against traded-book Sharpe for
positioning and the two funding variants, on the days positioning is
available; the labels give per-day churn. The two criteria rank the
three signals in opposite orders. (b) Sharpe and net return of the
funding z-score against holding period on the full sample, with the
weekly cell marked.

Two of the three objections hold, and one of them supplies a mechanism
this report had measured without explaining. The disagreement is
narrower than it appears: both accounts treat the funding rate as a
rent rather than a forecast, and differ on whether positioning collects
it better, which depends on whether the criterion is the coefficient or
the traded book.

## 12. The tradeable residue

Nothing in the six preceding sections adds a tradeable signal to the
funding rate. What is left is the carry, and the question becomes how
best to hold it. Figure 12 compares the strategies. Rebalancing every
three days and holding three days cuts costs from 4.8 to 1.8 bps per day
and smooths the path. The
z-scored funding decile swing variant nets $+8.9$ bps per day, $+32\%$
annualized, Sharpe 1.42, maximum drawdown $-40\%$, positive in six of seven
calendar years ($+84$, $+90$, $+37$, $+25$, $-38$, $+11$, $+52$; 2020 and
2026 are partial). A circular block bootstrap (15-day blocks, 1000
resamples) gives $p(\text{mean} \le 0) < 0.001$ with a 5% quantile of
$+16\%$ annualized; the focal-side decomposition shows both legs earn (long
excess $+5.7$ bps per day, $t = 1.6$; short excess $+5.8$, $t = 1.8$),
neither individually at conventional significance on that test, though the
same two cohorts clear a day-matched size-controlled placebo at $p = 0.005$;
the weaker verdict is the one carried forward. The additive composite on the
same 3-day schedule nets $+3.7$ bps per day (Sharpe 0.90, drawdown $-21\%$).

Two caveats belong next to those figures. The result is front-loaded: 2020
and 2021 supply roughly two thirds of cumulative net return, and 2020's
decile is two names against two on a median cross-section of 24. Excluding
those two years the variant nets $+3.7$ bps per day at Sharpe 0.77 ($t = 1.63$), which
does not clear a conventional threshold. And the swing schedule is not
demonstrably better than daily rebalancing: the same signal rebalanced daily
nets $+10.4$ bps per day at Sharpe 1.55, and paired on common days the
difference is $+1.5$ bps per day ($t = 0.61$, bootstrap $p = 0.27$). What the swing buys is a
shallower drawdown, $-39.5\%$ against $-44.8\%$, and under two fifths of the turnover
cost, 1.8 against 4.8 bps per day, which is what makes it the variant worth
carrying into a capacity discussion. It is not evidence that the signal
prefers a three-day horizon, and Section 6's term structure, in which
predictability is gone by twenty-four hours over the pooled sample, argues
against reading it that way.

![Figure 12](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig12_equity_curves.png)

**Figure 12.** Long-short funding strategies, net of 4 bps taker one-way:
raw-rate deciles with daily rebalance, 90-day z-score deciles with daily
rebalance, the additive composite on top and bottom 20% with a 3-day
swing, and the 90-day z-score deciles on a 3-day swing (gross shown
dashed). Log scale.

Robustness measurements (Figure 13): the information coefficient of the
raw rate decays from $+1.04\%$ at the 8h horizon to $+0.23\%$ at 24h and
turns negative beyond, which argues against a three-day hold rather than
for it and is taken up in Section 6; the
signal-embargo profile shows no isolated spike at zero lag, ruling out
look-ahead artifacts. Tails: worst day
$-10.1\%$, worst five days $-33.0\%$, return skew $+0.43$.

Every annualized Sharpe in this report scales daily volatility by the
square root of time, which assumes independent days. They are not: the
variance ratio at 21 days is 1.23 for the swing variant and 1.68 for the
daily raw-rate decile, so the quoted 1.42 and 2.40 deflate to 1.28 and
1.85. The ordering of the variants is unaffected because the correction
is similar across them, and the block-bootstrap p-values are computed on
means and so are unaffected entirely.

The drawdown figure understates how the equity curve is experienced. The
swing variant spends 85% of days more than 1% below its high water mark, 46%
more than 10% below and 22% more than 20% below, with a single unbroken
stretch of 904 of the 2203 days. At twice the gross notional, where the
return becomes interesting, those depths double.

The carry is an accrual at portfolio level and not per name. On the
short leg the realized 24h carry has the opposite sign to the rate at
entry 16% of the time, its 1st percentile is $-17$ bps and its worst
single name-day $-997$ bps; the long leg pays funding rather than
collecting it on 55% of name-days, its positive expectation coming from
a minority of deeply negative-funding names. The smooth portfolio carry
is diversification across roughly 36 positions, not a property of the
payoff. Liquidation and auto-deleveraging are not modelled anywhere: a
36-position book of alt perpetuals on one venue, at 23% annualized
volatility with a $-10\%$ worst day, carries margin risk for which the
per-name return clip does not stand in.

Capacity binds harder than a median suggests. An equal-weighted leg is
limited by its smallest member: that name has a median trailing volume
of USD 9.3M per day and 54% of leg-days contain one under USD 10M, which
puts 10% participation near USD 17M per leg rather than the USD 88M a
median-name calculation gives. Raising the liquidity floor, which is
what deploying size means, removes the return rather than scaling it. At
floors of 5, 25, 50 and 100M the price leg runs $+5.7$, $+4.0$, $+2.5$
and $-0.5$ bps per day and Sharpe falls 1.42, 1.02, 0.71, 0.35. The
carry is untouched at $+5.0$ to $+6.0$ but does not clear a threshold
alone. The price leg exists only in contracts too small to hold the
position, the same gradient Section 7 finds in the one-hour effect.

Cost sensitivity is measured rather than assumed, because the schedule is
a choice and the three variants churn at very different rates. Cost enters
the return linearly, so one run at a single basis point recovers the whole
curve exactly (Table 4). The swing variant pays 0.46 bps per day for each
basis point of one-way fee and would need 23.4 bps before its mean net
return reached zero, with net Sharpe falling to 1.0 only at 9.7 bps. Fees
are not what constrains it. The daily variant, churning 57% of each leg
per day, breaks even at 13.0 bps and the additive composite at 11.2 bps.
Across any schedule a reader is likely to face the three stay positive and
keep their order, which places the binding constraint on capacity and
drawdown rather than on cost.

**Table 4.** Net annualized Sharpe against the assumed one-way fee, and
the breakeven fee at which mean net return reaches zero. Cost is linear in
the fee, so the columns are exact rather than interpolated.

| Strategy | Gross | 2 bps | 4 bps | 6 bps | Breakeven |
|---|---|---|---|---|---|
| z-score deciles, 3-day swing | 1.71 | 1.56 | 1.42 | 1.27 | 23.4 bps |
| z-score deciles, daily rebalance | 2.25 | 1.90 | 1.56 | 1.21 | 13.0 bps |
| additive composite, 3-day swing | 1.41 | 1.16 | 0.90 | 0.65 | 11.2 bps |

![Figure 13](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig13_decay_embargo.png)

**Figure 13.** Robustness. (a) Information coefficient by forward horizon
with 95% intervals. (b) Embargo profile: z-score lagged k days against
the next-24h return, with no spike at zero lag.

## 13. Risk-managed carry

Every signal-side prescription tested above fails a control and the carry
survives, so the remaining question is whether the risk taken to collect it
can be reduced. Section 4 shows where the money goes: the funding leg
accrues almost without variance while the price leg taken on to collect it
draws down 43% on the swing variant. If that leg cannot be predicted, the
question is whether it can be made smaller per unit of funding collected.
Three standard carry constructions and a diversification test address that,
each against the same baseline with the paired test used throughout.

The first attempt is a trap (Figure 14a). Ranking on carry-to-risk, funding
divided by annualized volatility, produces an information coefficient of $+0.041$
($t = 12.9$), which would be the largest in this study. It is not a funding signal.
The inverse of volatility alone scores $+0.087$ ($t = 17.7$). Removing the volatility
rank linearly within each day still leaves $+0.025$ ($t = 8.8$), but a linear control
on a rank cannot absorb a nonlinear effect; demeaning both signal and
outcome inside volatility deciles within each day leaves $+0.003$ ($t = 1.3$). The
ratio was importing the low-volatility effect through its denominator, as
cross-venue dispersion imported the funding level in Section 10. As a
portfolio it is worse, not better: 6.4 bps per day at Sharpe 0.95 against
the baseline's 8.8 at 1.40 (this section needs a volatility estimate and so
runs on 418,464 symbol-days against Section 12's 423,216, which is why its
baseline differs in the second decimal), though it is positive in all seven
years where the baseline loses 40% in 2024.

![Figure 14](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig14_risk_managed.png)

**Figure 14.** Risk-managed carry. (a) The carry-to-risk ratio against
its components: the ratio scores an information coefficient of 4.1%, the
inverse of volatility alone scores 8.7%, a linear control on the
volatility rank leaves 2.5%, and demeaning inside volatility deciles
within each day leaves 0.3%. (b) Sharpe and mean return of the three-day
swing against long-short leg width; Sharpe peaks at quintile legs
(circled) while the mean falls monotonically.

Inverse-volatility weighting inside each leg changes nothing that clears
a threshold: Sharpe 1.46 against 1.40, a paired mean difference of
$-0.9$ bps per day ($t = -0.90$) and a bootstrapped Sharpe difference of
$+0.06$ ($p = 0.37$). Median annualized volatility is 99%, spanning 58%
to 177% between the 10th and 90th percentiles, so equal weighting does
put equal dollars behind unequal risk, but sizing it away does not
pay.

Market-factor neutralization fails because there is almost nothing to
neutralize. The decile long-short's beta to the equal-weight universe
return is $+0.013$, so the price leg's drawdown is idiosyncratic risk
rather than common-factor exposure. Hedging it with the universe basket
adds $+0.6$ bps per day ($t = 1.34$) and $+0.09$ of Sharpe
($p = 0.09$).

Idiosyncratic risk is not hedgeable but it is diversifiable, and widening
the legs prices that (Figure 14b). Sharpe traces an interior peak, 1.32,
1.40, 1.72, 1.32 and 1.15 at top-and-bottom 5%, 10%, 20%, 30% and 50%, while
the mean falls monotonically from 13.2 to 3.1 bps per day as the legs dilute
the funding spread they are built to harvest. Quintile legs of about 37
names are the best of these: Sharpe 1.72 against 1.40, at a paired mean
difference of $-1.4$ bps per day ($t = -0.87$) and a bootstrapped Sharpe difference of
$+0.32$ ($p = 0.084$). It trades mean for variance close to one for one, and at $p = 0.084$ it
does not clear a conventional threshold. Narrowing rather than widening wins
on the other metric: top and bottom 5% legs beat the decile by $+4.4$ bps per
day at $p = 0.073$, a smaller p-value than the quintile's Sharpe result, while
giving up 0.08 of Sharpe. The two criteria select opposite ends of the sweep
and neither clears a threshold, which is the finding.

## 14. Capture at the settlement timestamp

Section 13 leaves one lever untouched. Funding is charged only to positions
open at the settlement timestamp, so the same payment could in principle be
collected by holding two hours around it rather than three days, raising
carry per unit of price risk by an order of magnitude. This is the retail
funding-farming trade. It was tested on 1.73 million settlements across 670
contracts, opening at the hourly boundary before the timestamp, closing at
the one after, with the side taken from the previous settled rate so that
nothing unknown at entry is used. It does not work, and the reason is more
informative than the failure.

Costs are not the binding constraint (Figure 15). Among settlements whose
predicted rate exceeds the 8 bps taker round trip, 64% pay enough funding to
cover it outright, at a mean payment of 18.5 bps. What defeats the trade is
that the price moves against the party collecting by more than the payment,
wherever the payment is large enough to be worth collecting. Contracts
paying at least 8 bps to shorts see the price rise 24.7 bps over the
two-hour window against a mean payment of 14.9 bps. Around 99 contracts
settle at the same timestamp and their two-hour returns are one market move,
so the statistic is taken across timestamps rather than across settlements:
that gives $t = 3.4$, against $t = 9.5$ if the settlements are pooled as
independent. Contracts paying at least 8 bps to longs see the price fall
only 4.1 bps against a payment of 27.8 bps, and clustered that drift is not
significant ($t = -1.7$), so what the data supports is the asymmetry rather
than the universality. Run on the side known at entry, the price leg is
negative in seven of the eight side-and-threshold cells, and its magnitude
tracks the funding: collecting 55.7 bps on the long side at the widest
threshold costs 73.6 bps of price. Net of taker fees the trade loses 9.6 to
23.7 bps per settlement at every threshold, and the price offset exceeds
the payment before any fee is charged, so the result does not turn on the
schedule. The one positive cell rests on 976 events at $t = 0.93$.

![Figure 15](https://raw.githubusercontent.com/OctopusTakopi/funding-rate-alpha/main/results/figures/fig15_settlement_capture.png)

**Figure 15.** Capture at the settlement timestamp. (a) Funding received
by whoever collects it, against the two-hour price move running the
other way, by realized settlement rate; the offset grows with the
payment and exceeds it wherever longs are paying. (b) Net result per
settlement against the minimum predicted rate required to trade, on
taker and maker fee schedules, with the zero line for reference.

The settlement is public before it happens and the premium index
converges through the interval, so the price absorbs the payment.
Funding is compensation for holding the position, not free money on a
timestamp, and a two-hour window strips out the compensation along with
the risk. Timing the settlement does not separate the carry from the
price risk of an unhedged perpetual position, which is the construction
tested here.

## 15. Limitations

Costs are exchange fees only, quoted at a reference schedule of 4 bps
one-way taker; Table 4 gives the full fee curve, since the schedule is a
property of the account rather than of the data. No slippage, spread
crossing or market impact is modelled, so even the reduced capacity of Section 12, roughly
USD 17M per leg, assumes participation at no impact penalty. A funding
strategy is short the crowded side by construction, so the marginal
trade moves the rate it collects: true capacity is below the
volume-based number and the net figures are upper bounds.

Every strategy here is a directional cross-sectional book that holds
perpetual price risk. Delta-neutral cash-and-carry, long spot against
short perpetual, collects the same funding without that exposure, on a
different cost and capital base and against a different capacity limit.
It is outside this study and nothing measured here speaks to it.

The construction choices, a 90-day z-score window, a three-day holding
period, decile legs and a trailing USD 5M liquidity floor, are fixed
throughout rather than tuned on the sample. Only the composite of
Section 9 and the adaptive test of Section 6 are fitted on a training
window and evaluated forward; every other reported Sharpe, the swing
variant's included, is an in-sample estimate of a specification chosen
once.

The portfolio is dollar-neutral, not market-neutral, but Section 13
shows there is little to neutralize: the book's beta to the equal-weight
universe is $+0.013$ and hedging moves Sharpe from 1.40 to 1.41. The
price leg's drawdown, $-74\%$ on the raw level and $-50\%$ on the
z-score against $-29\%$ and $-45\%$ for the net strategies, is
idiosyncratic risk rather than factor exposure, which is why it cannot
be hedged away and can only be diversified.

Per-name forward returns are clipped to $[-90\%, +200\%]$ as a stop and
auto-deleverage proxy, since a squeeze in a small delisted contract is
otherwise unbounded. It is conservative rather than flattering: the cap
binds six times on each leg and removes more from the long side than the
short, so the clipped figures understate the strategy by 0.6 bps per day
on the baseline and a similar margin on the swing variant.

Six hypotheses are tested, several through more than one variant, without a
formal correction for multiple comparisons except in Sections 6 and 7, where
the twenty horizon tests and the twenty-four hourly tests each carry a
Benjamini-Hochberg correction: the venue basis of Section 10 clears a
two-sided threshold on its own but is dismissed on economic grounds rather
than statistical ones.

The tradable sample is the liquid tail, a trailing 90-day mean daily
quote volume of at least USD 5M, and the conclusions do not extend to
the illiquid majority. Open-interest analyses begin around 2021 because
causal top-150 membership requires a year of volume history. The
cross-venue tests assume execution on one venue and model no transfer,
margin or borrow cost of running a book across three, and the constant
interest-rate assumption affects the pinned-at-interest signature but
not the error result.

## 16. Conclusion

The practitioner note is correct that raw funding levels measure retail
orientation rather than crowding, and that per-contract rolling z-scores
with cross-sectional ranking are the right preprocessing. Its
cross-validation prescriptions are not. Short-horizon reversal is a real
standalone signal and open-interest change is largely the same signal
measured again; neither conditions high funding in a way that survives a
day-matched control, and no composite of them clears trading costs. Its
horizon prescription is neither confirmed nor refuted: returns are flat
across holding periods from one to five days and fall beyond, so the
three-day swing is a choice about turnover and drawdown rather than about
where the signal lives. The tradeable residue is carry harvesting, and it is
thinner than the full-sample figures suggest: two thirds of the return falls
in 2020-21 and the remainder does not clear a threshold.

The cross-venue prescriptions fail the same way. Venues genuinely
disagree about funding, by 33% to 38% annualized in the tail, but the
synthetic global index does not beat the venue a position is held on
once compared paired rather than by headline Sharpe, and open-interest
weighting makes it worse because it returns the dominant venue's own
rate at a correlation of 0.975. Cross-venue dispersion is a restatement
of the absolute funding level rather than an independent sentiment
signal, and it points the wrong way to trade. The pattern is consistent
across every claim tested: the preprocessing prescriptions are sound,
and every cross-validation prescription dissolves under a control that
holds the day, the funding level or the trading cost fixed.

The same holds on the risk side. The price risk that consumes the carry is
idiosyncratic rather than factor exposure, so neither carry-to-risk ranking,
inverse-volatility sizing nor factor hedging recovers a material part of it,
and leg width trades one against the other without either end of the sweep
clearing a threshold. What the funding rate offers this book is a carry
harvest whose capacity is set by how much idiosyncratic perpetual risk it
will hold.
The one signal from outside the note's framework that survives its controls,
the long/short account ratio of top traders, predicts at $+0.013$, below the
funding rate on the coefficient and above it on the traded portfolio, its
advantage being turnover rather than accuracy.

On the question itself, funding does carry information about future prices,
in the level rather than the z-score, at horizons below eight hours, and in
the continuation direction. Pooled it is worth less than a 4 bps one-way
round trip, though from 2024 onward it clears one at any competitive
tier, and it is a measurement of the basis rather than a forecast of
value. What remains
tradeable is not prediction but carry, and the preprocessing the source note
prescribes earns its place by stabilising that carry across regimes rather
than by forecasting anything.

## Reproduce

Two things must be in place first. `FUNDING_DATA_ROOT` names the archive
root and every stage reads it; without it the scripts look in
`~/binance_data` and fail on the first stage. And the pipeline expects
hourly klines already parsed to `$FUNDING_DATA_ROOT/parquet_1h/`, one
parquet per contract with columns `open_time, open, high, low, close,
volume, quote_volume`; that conversion is upstream of this repository.

```
export FUNDING_DATA_ROOT=/path/to/binance_data

make universe manifests parse-funding download-metrics download-premium \
   download-premium1h parse-metrics panel research
```

`make research` runs the single-venue analysis and renders its figures. The
cross-venue stages are separate, for the reason given below. Figure 1 and
`make premium-check` need the 1h premium archives from `download-premium1h`;
without them the figure is skipped with a warning and the check reports its
result as inconclusive.

The cross-venue extension of Section 10 has its own chain, fetching the
other two venues' archives and reproducing Figures 9 and 10:

```
make cross-venue-all
```

Download stages are resumable, skipping any archive already present.
Sources, coverage floors and the survivorship construction are
documented in `docs/venues.md`. The archives are public; this repository
carries the code, the frozen manifests and the text and figure outputs,
not the data.

Released under the MIT License. This is a study of historical data, not
investment advice; Section 15 sets out what the results do not
establish.

funding-rate-alpha

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README