From a measurement to a rule: what the open index can do on its own, and what it cannot

1. The question

The series so far has measured. A headline describes a move and predicts nothing (note 1, note 6); the flow of headlines about a coin, measured against the coin's own norm, is followed by about 20 bps a day of abnormal return for two days (notes 1 and 5); the volume of coverage predicts the size of moves and the tone their direction (note 2); and a naive long-short sort on the index pays its fees only when the book is held longer than the signal and sized small (note 8). A reader who wants to use the index rather than study it will now take five further steps, and each is a choice a practitioner makes by habit: trade only strong signals, drop the short leg, wait for a better price, size by liquidity, or do none of that and let the index tilt a book that exists for other reasons. This note takes the five steps on the open index and reports what each is worth, with the same accounting as note 8. Three of the five do not survive it, and the reason each fails is more useful than the two that pass.

One thing is deliberately outside the note. The live account described in note 3 trades a rule built on the same data with a different construction of the index and a different entry and exit. Nothing here is that rule, nothing here is fitted to it, and the note does not claim that what works or fails on the open index explains it. Its test is the live record.

2. Data

Every scored headline from 4 October 2024 to 23 October 2026 about one of the 37 coins with a Binance USDT spot pair and at least 500 headlines, 211,663 headlines, aggregated into the 24-hour index of note 5: the mean score of a coin's headlines in a calendar day with at least three of them, expressed as a z-score against the coin's previous 30 observed index days with at least 20 of them. 9,805 coin-days carry an index; 729 daily rebalances from 25 October 2024 to 23 October 2026, of which 668 have at least eight coins with an index. Prices are 30-minute Binance closes; returns are log returns, abnormal against Bitcoin for the long-short constructions, raw for the long-only ones, which hold no short leg to net against. Costs follow note 8: turnover is the notional traded, $T_b = \sum_c |w_{c,b} - w_{c,b-1}|$, cost is $T_b$ times the fee per side, reported at 0, 5 and 10 bps per side, and the impact model is eq. 3 of note 8. The long-short rule with the |z| ≥ 1 filter reproduces note 8's numbers to the second decimal (table_consistency.csv): the two notes share one code path.

3. The five steps

Strength. The long-short sort of note 8, thirds, held until the coin leaves its group, with the coin required to have $z \ge \theta$ to enter the long leg and $z \le -\theta$ for the short, $\theta \in \{0, 0.5, 1, 1.5, 2\}$. The same for the long-only rule below.

Long only. Long the top third (and separately the top fifth) of coins with an index, equal weights, $\sum w = 1$, held until the coin leaves the group, or rebuilt daily, with the same filter grid. Two benchmarks on the same 729 days: Bitcoin held throughout, and the equal-weight basket of every coin with an index that day. The number that isolates the index's selection from the market is the rule's excess over that basket on the days it is invested.

Dip entry. For every opening of the long-only top-third rule with |z| ≥ 1, 859 signals, the entry at the first 30-minute close after the window end, $P_0$, is compared with waiting up to $W$ hours for a 30-minute close at least $d$ below $P_0$ and then buying at the first later close that is above its predecessor, the turn; $d \in \{1\%, 2\%, 3\%\}$, $W \in \{24, 36\}$ hours. A signal that does not dip within $W$ is skipped. Exit is the same for all: the coin leaves the top third, or seven days from the signal, whichever first.

Sizing. For the long-only |z| ≥ 1 rule, equal weights against weights proportional to the coin's median 30-minute dollar volume, weights inversely proportional to one plus the coin's attention z-score of note 2 where it is positive, and both, each capped at three times the equal weight. Capacity is the book size at which the net return at 10 bps falls to zero under the impact model.

Overlay. An equal-weight long-only book of the ten largest coins by dollar volume, rebalanced daily, is the reference; it earns $-1.6$ bps a day over the sample. Tilt: $w_i \propto 1 + \kappa z_i$ with $z$ clipped to $[-2, 2]$, coins without an index at $z = 0$, renormalised, $\kappa \in \{0.25, 0.5, 1\}$. Timing: the reference book, but a coin with $z \le -1$ that day is not held and its weight goes to the others. For each: the tilt's contribution over the reference, its Newey–West $t$, the extra turnover, net contribution at 5 and 10 bps, tracking error, and impact at book sizes of 1, 10 and 50 million dollars with each trade's notional equal to the book times the weight change.

4. Results

4.1 The threshold is a spike

Table 1. The strength filter. Basis points per day averaged over all 729 days, flat days counting as zero; net at 10 bps per side; t is Newey–West. Bold: the threshold of note 8.

Note 8 reported the filtered until-exit rule at 35 bps a day gross and 19 net and said the number was the best of 84 constructions. Fig. 1 shows what that means. The threshold of 1 is a spike: at 0.5 the same rule earns 19.5 gross and 3 net, at 1.5 it earns 27 and 18 on a book that is invested 38% of the time, and at 2 it earns nothing. A filter that worked as a principle would trace a hill, rising as the weak signals are dropped and falling only when the book becomes too thin to diversify. This one jumps. The long-only line is the same shape with lower values: 1 bps a day without a filter, 19 at the threshold, 1 again beyond it. The honest figure for the filtered long-short rule is therefore not 35 but the range of its neighbours, 20 to 27 gross and 3 to 18 net, and the honest figure for the unfiltered rule is the one in the first row, 22 bps gross and 7 net, which is what note 8's Table 3 already showed. The filter concentrates the book, 73 openings a month instead of 141, and the concentration happens to pay at one setting on this sample.

4.2 Long only: the right coins, in the wrong years

Table 2. Long-only rules, raw returns, Σw = 1, bps per day over all 729 days; "excess" is the rule minus the equal-weight basket of all coins with an index, on the days the rule is invested, with its Newey–West t. The full grid of thresholds and breadths is table_long_only.csv.

Two numbers in Table 2 are true and point in opposite directions. The excess over the basket is the index doing what the earlier notes said it does: on the days the filtered rule holds coins, those coins beat the average coin with news by 17 bps a day, $t = 2.3$, and the unfiltered rule by 6. The index picks the right coins. The level is another matter. The filtered rule's $t$ on its own gross return is 1.3, its cumulative net return of 71% is in Fig. 4 a staircase whose first step is the whole climb, and the subsample rows below say why.

Table 3. The filtered rules by subsample. The last quarter of 2024 is 68 days of the 729.

The long-only rule earned 86 percentage points of its 71% in the last 68 days of 2024, when the basket itself rose 65 bps a day and Bitcoin 47, and has lost money net since: 6.5 bps a day gross from January 2025, $-2.3$ net, $t$ indistinguishable from zero. The long-short rule holds up better, 29 bps gross and 12 net since 2025 with a $t$ of 2.9 on the gross, because its short leg earns in the quarters the long leg does not; but 60 of its 140 points are also the first 68 days. A long-only rule on this index is a bet that the coins the news favours rise, and in a market that falls for most of a year they fall less, which is what the excess measures and what a long-only book cannot spend.

4.3 The dip selects the losers

Table 4. Entry on a dip. Returns in basis points per trade, net of 10 bps per side where stated; "same trades from P0" is what the filled trades would have returned entered at the close instead; "skipped signals from P0" is what the signals that never dipped returned when entered at the close. Exit is the same throughout.

Waiting for a better price is the oldest habit in trading, and on this signal it is the most expensive. Every variant fills fewer signals and loses on the ones it fills: $-35$ to $-56$ bps a trade net, against $+2$ for buying at the close. The two right-hand columns explain it. The trades that do dip would have lost 180 to 377 bps had they been bought at the close, so the dip buyer does get a discount of 1.4 to 3.5 percent, and still loses, because the coin keeps going; the signals that never dip, the ones the rule skips, are where the return is, 210 to 391 bps a trade. On the open index a dip after a positive signal is not an entry; it is the signal failing. The rule that waits for it systematically keeps the failures and discards the successes, and no width of dip or length of wait in the grid changes the sign.

This result belongs to the open index with a mechanical exit, and what it says is that a dip is not an entry on its own. A rule that waits for one needs its own answer to the question of which dips, and the answer is the construction of the signal and the rule of exit: for instance, whether the news that produced the signal is still there while the price falls, or has already turned. The live account of note 3 gives that answer with its own construction of the index, an entry that is cancelled when the signal reverses, and an exit on the signal; its filled entries, bought a median 1.6 to 1.8 percent below the signal price, match the simulator in 92% of cases. What that construction does with a dip is measured there, trade by trade, and not here.

4.4 Sizing by liquidity doubles the capacity

Table 5. The long-only |z| ≥ 1 until-exit rule under four sizing rules, bps per day; impact by the model of note 8 with each position's notional equal to the book times its weight. The cap binds on a fifth of invested days; with about three coins held, volume sizing mostly means the most liquid of them carries most of the book.

Sizing changes the capacity, not the edge. Weighting by dollar volume raises the gross from 18.6 to 23.9 bps a day, mostly because the liquid coins happened to do well on this sample, and halves the impact per unit of return, so the book size at which the rule stops paying its fees moves from 126 to 258 thousand dollars. Sizing against attention, the note 2 idea that loud coins move more and should be held smaller, costs a little return and a little capacity; the two combined land between. None of this changes the order of magnitude of note 8: six figures.

4.5 Overlay: the only thing that pays is not buying into bad news

Table 6. Overlays on the equal-weight book of BTC, ETH, SOL, XRP, DOGE, BNB, SUI, PEPE, TRX and ADA, which itself earns −1.6 bps a day over the sample. Contribution is tilted minus reference, gross; tracking error is the annualised standard deviation of the daily difference; impact per day by the model of note 8 on the tilt's trades alone.

For a book that holds the large coins for other reasons, the index is worth little as a tilt and something as a veto. Tilting weights toward the coins with the higher index adds 0.6 to 2 bps a day gross, with a $t$ of about 1 and a tracking error that grows faster than the return; at 5 bps per side the tilt is flat, at 10 it loses. Declining to hold a coin on the days its index is below $-1$, which happens on 57% of days to 1.3 coins on average, adds 2.8 bps a day with $t = 2.2$, keeps 1.2 net at 5 bps, and loses half a basis point at 10. At ten million dollars the impact of either is a few basis points a day, at fifty million it is more than the contribution. The arithmetic is the same as in note 8 from the other side: the index moves the return of large coins by a few basis points a day, which is more than nothing and less than the cost of moving a large book every day.

4.6 Execution

Delaying the long-only rule's entries by half an hour, an hour, two or six hours changes its gross between 17.9 and 22.4 bps a day with no pattern; the long-short rule, as in note 8, loses a fifth of its return by six hours. A rule with three positions and a $t$ of 1.3 does not resolve a lag; the signal is known at the window end and the earlier notes say the first hours matter for the two-sided book.

5. Robustness

Table 7. By calendar quarter, bps per day; the last column is 23 days of October 2026.

The long-only rule's quarters track the basket's with a positive offset: when the market falls, it falls less, and it never earns against the market. The filtered long-short rule is positive gross in eight quarters of nine and positive net in five; the unfiltered one in six. October 2026, 23 days, is negative for every rule in this note and for the basket, as it was for everything in note 8 and for the live account in note 3. The definitions the constructions leave open, the odd coin of a cross-section, the norm window and the price-bar convention, are the ones note 8's Table 7 varies; this note fixes them as note 8's main specification does and lists each in the published definitions file.

6. What this says, and what it does not

Of the five steps a practitioner takes from a measurement to a rule, the open index supports two and refuses three. It supports selection: on the days it names coins, those coins beat the average coin with news by 17 bps a day, and a large book that simply declines to hold the coins it marks worst collects 2.8 of them with a $t$ above 2. It refuses the rest. The strength threshold that made note 8's best construction is one point on a spiky line, and the rule's honest return is its neighbours'. A long-only book built on it has the right coins and the wrong years: its level is a single quarter, and since then it has earned nothing net. And the entry habit most traders would add first, wait for the dip, inverts the rule, because on this signal the dip is the failure and the signals that do not dip are the ones that pay.

For a user of the data the practical content is in the two that pass and the manner of the three that fail. As an input, the index says which coins to prefer and which days to stay out, and it says so with enough margin to survive maker fees on a large book that is trading anyway. As a rule on its own it is a six-figure, regime-dependent opportunity, which is what note 8 said from the cost side. A rule that does more than that must bring its own construction to the data, and the open index cannot vouch for it; what vouches for it is a record kept against a simulator in public, trade by trade, which is the design of note 3, and the reason that note is the one in this series that will keep being updated.

What this note does not do: it does not test constructions of the index other than the open one, it does not model a short leg's funding, and its impact model is the coarse one of note 8. All of its tables and the code paths behind them are in the published files, and a reader with a construction of their own can run the five steps on it.