What it would take to trade the flow: turnover, costs and holding period

1. The question

The first note measured what a coin's news tone, taken against its own norm, knows about its next days: about 25 bps a day of abnormal return for two days, in a portfolio long the third of coins with the highest index and short the third with the lowest. It also said, in one sentence, that this is not tradable as published: the portfolio is rebuilt every twelve hours and a round trip on a spot exchange costs more than the portfolio earns between two rebalances. This note takes that sentence apart. The gross return of a sort is a property of the information; its cost is a property of how the sort is built. Between the two there is a map, and the question is where on it the information survives.

Three things are held fixed. The signal is the index of the earlier notes, unchanged. The universe is the 37 coins of the first note. And the configuration that is actually traded by the live account described in note 3 is not among the constructions shown: it uses the same index but a different entry rule, and this note is about the generic sort, not about it. What the live account does contribute is 57 real fills, which calibrate the slippage model in Section 3.3.

2. Data

Every scored headline from 4 October 2024 to 23 October 2026 about one of the 37 coins, 211,663 headlines and 485,042 coin–headline rows, aggregated into the index of note 5: the mean score of a coin's headlines in a window of 12 or 24 hours, at least three headlines, expressed as a z-score against the coin's own previous 30 observed index windows (60 for the 12-hour index), at least 20 of them; a window is observed when it has three headlines, so in thin months the norm reaches back further than 30 calendar days. Prices are Binance USDT spot closes at 30-minute resolution through 24 October; all returns are log returns and the abnormal return of a coin is its return minus Bitcoin's return over the same interval; Bitcoin itself enters with its raw return, as in the earlier notes. The 12-hour index exists for 14,748 coin–windows and the 24-hour index for 9,805 coin–days.

Table 1. The four rebalancing grids. The 48-hour and weekly grids sample the 24-hour index on alternate days and on Mondays at 00:00 UTC; the weekly grid has too few rebalances to measure anything and is reported for completeness only.

3. Method

3.1 Constructions

A construction is a rule that turns the index at a rebalance time $e_b$ into a vector of weights $w_{c,b}$ over coins, with $\sum_c w_{c,b} = 0$ and $\sum_c |w_{c,b}| = 2$: one unit of capital long, one short, equal weights within each leg. Four choices define it.

Frequency. The portfolio is reformed every 12 hours, 24 hours, 48 hours or 7 days.

Breadth. Long the top third and short the bottom third of coins with a signal, as in the first note, or the top and bottom fifth. A rebalance needs at least eight coins with a signal for thirds and ten for fifths; otherwise the book is flat for that interval and the interval counts as zero in every average below.

Holding rule. (a) One interval: the book is rebuilt from scratch at every rebalance. (b) Fixed horizon $k$ longer than the interval: at each rebalance a new vintage is formed and held for $k$ intervals, and the book is the equal-weighted average of those vintages formed in the last $k$ intervals that exist, in the manner of Jegadeesh and Titman. A rebalance without a tradable cross-section, or one at which the filter empties a leg, forms no vintage and its share goes to the live ones, so the book is fully invested whenever at least one vintage is live; it is not a book that holds $1/k$ in cash for every missing vintage, and its coverage, turnover and coins per leg below are those of the fully invested average. (c) Until exit: at every rebalance with a tradable cross-section the coins with a fresh index are ranked as in (a); a coin enters a leg when that ranking puts it in the top (bottom) group, and a held coin leaves when it has a fresh index at a rebalance and that same ranking puts it outside its group (with a strength filter the group is the filtered one, so a held coin whose fresh z-score has fallen below the threshold leaves). A held coin with no fresh index at a rebalance is not ranked and not tested, and stays where it is; it is closed after seven days without any index. At a rebalance without a tradable cross-section nothing enters or leaves except by the seven-day rule. The legs are renormalised to equal notional at every rebalance, and the book is flat while either leg is empty. A strict variant that exits as soon as the coin is not ranked in the group is reported in the robustness table.

Strength filter. None, or a coin enters a leg only if its own z-score is at least 0.5 or 1.0 in absolute value with the right sign. The filter reduces the cross-section and therefore coverage; both are reported.

With four frequencies, two breadths, up to five holding rules and three filter levels that is 84 constructions.

3.2 Returns, turnover and break-even

Gross return per rebalance is $\sum_c w_{c,b}\, AR_{c,b}$, where $AR$ is the coin's abnormal return from the first 30-minute close at or after $e_b$ to the next rebalance, averaged over all rebalances of the grid including the flat ones, with a Newey–West standard error whose lag is the holding length in intervals. Turnover is measured in the only unit that matters for cost, the notional actually traded:

where $c$ is the one-way cost in basis points of notional. Since $\sum|w| = 2$, a full replacement of the book has $T_b = 4$; the tables also report $T_b/4$, the share of the gross book replaced. The break-even fee is the one-way cost at which the construction earns nothing:

a single number in basis points per side that can be set against any exchange's fee schedule. Net returns are reported at $c \in \{2.5, 5, 7.5, 10\}$ bps per side: 10 is the base spot taker and maker tier on the largest exchange, 7.5 the same with the exchange-token discount, 2.5 to 5 the VIP tiers that need monthly volume in the tens to hundreds of millions.

3.3 Spread, impact and capacity

Fees are the known part of the cost. The rest is estimated from the 30-minute bars, per coin. The effective half-spread is the Corwin–Schultz estimator on consecutive bars, which uses the fact that the high–low range over two bars contains the spread twice but the variance only once; at 30-minute resolution on liquid pairs the estimator is noisy and its negative draws are floored at zero, which biases it upwards, so both the floored and the unfloored average are reported and treated as an upper and a lower bound. Price impact for a position of notional $N$ is a square-root rule,

with $\sigma_{\text{bar}}$ the coin's median 30-minute log range and $V_{\text{bar}}$ its median 30-minute dollar volume at rebalance times, and the coefficient one half an assumption in the range the microstructure literature reports. Total one-way cost is fee plus half-spread plus impact at $N$ of 10 thousand, 100 thousand and 1 million dollars per coin position. Capacity is read off the same model in two ways: per coin, the notional $N^{*}$ at which impact equals the coin's mean signed abnormal return when it sits in a leg of the 24-hour sort; and for the book, the gross notional at which the impact of the daily rebalancing alone equals the gross return.

The model is checked against reality in one place. The live account of note 3 enters positions with a pending order that waits for a dip and fills on a real price; its 57 matched fills of the last eight weeks can be compared with the simulator's price for the same entry. The median difference is 6 bps in the account's disfavour, with quartiles at $-19$ and $+24$ bps, and the exit prices agree to within half a basis point. For positions of the size that account trades, realised spread plus impact is therefore single-digit basis points, which is the unfloored Corwin–Schultz column and not the floored one.

4. Results

4.1 The two baselines

Table 2. The sort of the first note, long the top third and short the bottom third, rebuilt every interval, no filter. Gross and net in basis points per rebalance averaged over all rebalances of the grid, flat intervals counted as zero. The 12-hour sort has a book in 68% of intervals because the noon window rarely has eight coins with three headlines.

The 12-hour sort earns 7 bps per rebalance on average and trades 1.9 units of notional to do it: nearly half the book is replaced twice a day, 351 times a year. It breaks even at 3.7 bps per side, a fee available to almost nobody, and at the base tier it loses 12 bps per rebalance, 88% a year, with a $t$-statistic of $-2.9$. This is the sentence of the first note, with numbers. The 24-hour sort is a different object: it earns three times as much per rebalance, 21 bps, and trades only a quarter more, so its break-even is about 8.5 bps per side. At 10 bps it loses about 4 bps a day, indistinguishable from zero; at 7.5, the fee with the exchange-token discount, it earns 2 or 3; at 5 about 9. These are the figures of one computation, and Section 5 shows how far choices the construction leaves open can move them. Half the gap between the two grids is coverage and half is horizon: the first note already found that the index predicts two days of return, and a book held for one day collects more of that than a book held for half a day.

One component of the 21 bps is not cross-sectional. Because Bitcoin enters with its raw return (Section 2), the book is long or short Bitcoin outright on the days BTC sits in a leg, and on those days it collects Bitcoin's market return. Without BTC's own return the 24-hour baseline earns 18.7 bps a day, breaks even at 7.5 bps and loses 6 bps a day at 10 bps; about 9% of the gross in Table 2 is Bitcoin's market return, and the book is not Bitcoin-neutral when BTC is in a leg. Turnover, coverage and coins per leg are the same under both accountings.

4.2 The map

Fig. 1 is the note. Reading it by rows: the one-interval rule has the lowest break-even at every frequency, because it replaces most of the book every time; a fixed horizon of 48 hours or 7 days with overlapping vintages roughly doubles the break-even at the same or slightly lower gross, because each vintage trades once in and once out; and holding until the coin leaves its group does the same by a different route, leaving in place coins that have no fresh signal to rank them out. Reading it by columns: the 12-hour frequency is pale everywhere, and the strength filter makes it worse, not better, because at 12 hours a coin with $|z| \ge 1$ is rare and the book is flat most of the time. The dark cells are all at 24 and 48 hours.

Table 3. A selection of the 84 constructions; the full set is table_constructions.csv. Gross and net per day (12-hour and 48-hour figures converted), coverage is the share of rebalances with a book, coins per leg long / short, $T_b$ the notional traded per rebalance in units of one leg. Bold: the highest break-even and the highest net at 10 bps among constructions with at least 100 traded rebalances. Standard errors are Newey–West with the holding length as lag.

Of the 84 constructions, 45 have a positive net return at 10 bps per side: 27% of those rebalanced every 12 hours, 88% of those rebalanced daily, 72% at 48 hours, and the weekly grid is not measurable. The two levers that matter are visible in the table. Holding longer than the signal interval, by overlapping vintages or by waiting for the coin to leave its group, cuts the notional traded by 40 to 80% at a gross return that is unchanged or higher, because the index's information lasts two days and the second day was being thrown away. The strength filter at $|z| \ge 1$ concentrates the book in two to three coins per leg and raises the gross to 35 bps a day when combined with the until-exit rule, 36 in the fifths version, with a $t$ of 3.5 and a break-even of 22 to 25 bps. Those two figures also depend on a detail of the exit rule that the data do not fix: under the alternative reading in which a held coin without a fresh index is re-ranked on its last z-score for the exit test, the filtered until-exit constructions earn 30 bps a day gross and 16 net at 10 bps ($t$ = 1.5), so the honest range for the filtered rules is 13 to 21 bps a day net at the base fee, with the upper end resting on that reading. The highest break-even belongs to the least active rule, a 7-day overlap with the filter, which trades two thirds of a unit of notional per day and breaks even at 30 bps.

Two cautions belong next to these numbers, and the second is the larger. First, the best constructions are the best of 84 tried on one sample, and the within-day permutation test, which shuffles the index across coins 500 times and finds the real net return of the two until-exit constructions outside every draw ($p < 0.002$ and $p = 0.002$), answers only whether the index is related to return within a day; it does not correct for choosing the construction after seeing the result. The honest reading of Table 3 is not the top row but the shape: lower turnover at constant gross, everywhere in the 24-hour column. Second, fees are the smaller half of the cost for a sort of this kind, and the next section is about the other half.

4.3 Spread, impact and what is left

Table 4. One-way costs in basis points of notional from the 30-minute bars, eq. 3 for impact. The floored Corwin–Schultz estimate overstates the spread of liquid pairs by an order of magnitude against the live fills of Section 3.3; the unfloored column is the one consistent with them. Full table: table_slippage.csv.

The impact column is where the book meets the market. For BTC a position of a million dollars costs 3 bps of impact; for the median coin of the universe it costs 66, three times the gross return of the baseline. The 24-hour baseline trades 2.5 units of notional a day, so with the lower-bound spread alone, before any fee, it pays about 7 bps a day in spread; with the upper bound it pays 28 and is under water at zero fee. The constructions that trade little fare better in proportion: the 7-day overlap with the filter pays 8 bps a day of upper-bound spread and, at ten thousand dollars per position, 3 of impact, and keeps 9 bps a day at zero fee and 2 at 10 bps. At a hundred thousand per position its impact rises to 11 and the net return is zero at any fee; at a million it is $-28$ bps a day.

Capacity turns the same model around. Per coin, the notional at which impact eats the coin's own contribution is 7.9 million dollars for BTC and 5.7 for XRP, the two coins with the deepest books and the smallest per-coin return, between 0.2 and 2.7 million for a middle group, and in the thousands for the smallest coins; summed over the 25 coins with a positive contribution it is 27 million dollars. Twelve coins, ETH, BNB, SUI and UNI among them, have a negative mean signed return when in a leg and no capacity at all. For the book as a whole, the gross notional at which daily rebalancing impact alone equals the 21 bps of gross return is 285 thousand dollars at zero fee, 142 thousand at 2.5 bps, 48 thousand at 5, and zero at 10, because at 10 bps the baseline is already under water before impact. The three tiers of Table 5 say why: of the 21 bps, BTC and ETH contribute 0.9, the next eight coins by volume 4.1, and the remaining 27 coins 16.0. The first figure is Bitcoin's alone: ETH's contribution is negative, and BTC's is its raw return, which is how it enters the book (Section 2) and why it also has the largest per-coin capacity in Fig. 3. The return lives where the books are thin.

Table 5. Where the 24-hour baseline's return comes from. Contributions are the sum over the tier's coins of weight times abnormal return (for BTC, its raw return) and add up to the baseline gross; the within-tier sort rebuilds the thirds inside the tier alone, which needs at least six coins.

4.4 Execution lag

Table 6. The baselines with the entry price taken later than the first close after the window end, bps per rebalance.

The index is known at the end of its window, and a book that trades two hours later gives up a fifth of the daily return, 21 to 16.5 bps, and a fifth of its break-even. The first half hour costs 6%. The 12-hour baseline shows no decay, which is consistent with the first note's finding that the 12-hour index is the noisier measurement of a two-day effect. The collection delay of the dataset itself, 11 minutes in the median from publication to the database, is inside the first bar.

5. Robustness

Table 7. Variants of the main constructions; the BTC-neutral rows are measured against the baseline without Bitcoin's own return (Section 4.1), each changing one further choice; full set including the weekly phases in table_robustness.csv.

Without BTC and ETH the 24-hour baseline is unchanged, 21.6 bps and a break-even of 9.6, and the 12-hour baseline disappears, 1.6 bps, which repeats the first note. The until-exit rule keeps its advantage, 17 bps of break-even. The strict variant of that rule, which exits a coin as soon as it is not ranked in its group at a rebalance rather than when it is ranked outside it, has the baseline's turnover and the baseline's break-even: the whole saving comes from leaving alone coins that have no fresh signal, not from any property of the exit. The 48-hour grid depends on which days it falls on, 15 bps of gross on even days against 21 on odd, which is one reason the 48-hour and weekly columns of Fig. 1 should be read as noisier than the daily one.

Table 8. By calendar quarter, bps per day; the last column is 23 days of October 2026. The until-exit row uses the rule without a filter.

The gross return of the baseline is positive in eight full quarters of eight, but it is 64 bps a day in the last quarter of 2024 and between 3 and 35 in every full quarter since, so the break-even moves between 1 and 27 bps per side across quarters and the net return at the base fee is positive in three of them. The until-exit rule is positive in six of eight. The 23 days of October 2026 are negative for both, gross as well as net, which is the sample-size warning of note 3 in another form. Fig. 4 shows the same thing in one picture: the baseline gross climbs steadily, the baseline net at 10 bps earned everything it would ever earn by February 2025 and gave it back, and the construction that survives does so with a flat eighteen months in the middle.

A last set of variants concerns choices the construction does not pin down rather than the data. Holding everything else fixed, the gross return of the 24-hour baseline moves between about 15 and 25 bps a day depending on which leg receives the odd coin of a cross-section that does not divide by three, on the length of the norm window (20, 30 or 60 observed windows), on whether windows with one or two headlines enter the norm, and on the price-bar convention at the window end; the BTC-neutral rows of Table 7 give the figures. Coverage, turnover and the number of coins per leg do not move at all across these variants, so each is the same book with a slightly different ranking inside it. The conclusion of Section 4.1 holds in every one of them: the daily sort sits near break-even at 7.5–10 bps per side, with the break-even fee itself between 7 and 10 bps. The decimal places of Table 2 should be read in that light; they are the precision of one computation, not of the quantity.

6. What this says, and what it does not

The information in the flow is what the earlier notes said it was: a cross-sectional sort on the news index earns about 20 bps a day gross, at daily rebalancing, with a $t$ between 2.5 and 3.7 depending on construction, and the effect lasts about two days. What this note adds is the price of collecting it. Rebuilt twice a day it cannot be collected at any fee a trader can get. Rebuilt daily it covers a base-tier fee only if the book is held longer than the signal interval, and the rules that do that raise the break-even from 8.5 to between 15 and 30 bps per side; with the strength filter they leave 13 to 21 bps a day at the base fee, numbers that carry a selection caveat, a $t$ of around 2 and, at the upper end, a dependence on how the exit rule treats a coin without a fresh index. Spread and impact are the larger cost, and they bound the size: a few hundred thousand dollars of gross notional for the book, ten thousand per position for the construction that keeps the most, with the return concentrated in the 27 coins whose books are thinnest and absent in ETH. Two hours of delay cost a fifth of it.

That describes a particular kind of opportunity. As a strategy on its own it is real, it is small, and it belongs to a book that trades at maker fees, in size that does not move a thin pair, within an hour of the window closing. The capacity bound is a bound on that naive sort: a book that holds the 37 coins for other reasons and lets the index tilt its weights, time its entries or size its risk is not rebalancing a long-short portfolio every day, and nothing in this note limits what the index is worth to it. For a user of the data the conclusion is the one the first note reached by a shorter road: the value of the index is as a measurement of what the market is being told and as an input to models that trade for other reasons, and the next note is about how those inputs are combined into a working rule. The configuration of the live account in note 3 is a different construction on the same index, sized for exactly the conditions described here, and its results are reported there as they come.

Two things this note does not do. It does not model funding or borrowing for the short leg, which on spot is not available at all and on perpetuals carries a funding rate that varies with the same sentiment the index measures; a long-only version of every construction is a separate question. And its spread model is coarse: 30-minute bars are the resolution of this dataset, and a reader with order-book data can replace Table 4 with measured numbers and rerun the rest from the published files.