Two cuts in 122 days, and what they cost

Dear Partners,

Letter No. 1 ended on an open position. We wrote that we did not know how it would end, and that we would have written the letter differently if we had. Now we know. It ended in a second cut.

The short version: between 1 June and 1 October the account went from 50 ETH to 45.10 ETH (−9.8%), while ETH itself rose 34%. This is exactly the scenario we described in the third part of the last letter as the price of the structure. We asked you to remember that in a rally we would look foolish. The market took the request seriously.

What follows, in order: the account, what we changed in the rules, what we tested and threw away, where we got things wrong ourselves, and what all of it costs.

Starting capital is still 50 ETH, deposited on 1 June, with no top-ups. Period: 122 days.

Two rows should be read together. The stack (45.10) counts only what is closed. The open position currently carries an unrealised -2.52 ETH, so with it the account would be worth about 42.6 ETH (−14.8%). We show both, because on the date of this letter the second is the more honest number.

And a second pair: in coin we lost almost 10%, in dollars we made almost 21%. Both are true and neither is a victory: simply holding would have returned 0% in coin and +33.9% in dollars.

Act four, 9 August to 4 September: the second position, and this was the test. Entry of 6248 contracts at 1916.17, target 1849.10. The pullback never came: ETH went from 1916 to 2504. On day 26 the circuit breaker did what it did the first time: it closed the position at a loss of -7.663 ETH. At its worst point the price stood 33.9% above our entry.

Two cuts in a row, -7.202 and -7.663, -14.865 ETH together. The eight June wins brought in +9.616. The whole arithmetic of the account fits in one line: many small wins, and each large loss eats all of them.

Act five, 4 to 18 September: two quiet weeks. No trades. The rule forbade entering a confirmed rally, and in those two weeks it blocked 55 entry slots. Together with the first pause (about 70 slots in July and August), the breaker and the gate have kept us out of the market for 125 slots since the account began. How much money that saved we do not know and cannot know: there is no way to measure the loss on a trade that did not happen.

Act six, from 18 September: the third position. Entry of 6820 contracts at 2441.22. Three days later, on 21 September, ETH rose to 2806, which is 15% above our entry, and that is exactly where the account's deepest point sits: a 31.0% drawdown from the peak. On 1 October the position is open and the price is around 2683. The rule gives it 26 days and 13 have passed. As last time, we are writing this letter without knowing the ending.

The win rate got lower (80% against 89%), and that is the only thing that changed in it. The average win is 1.2 ETH, the average loss 7.4 ETH, more than six times as much. If someone has quoted you a percentage of winning trades without saying the next sentence, assume you were shown half a page.

In the last letter we wrote that drawdowns would be deeper than the 13.09% of the first period. They are 2.4 times deeper. The nine-year measurement gave 34.8%, and four months were enough to reach 31.0%.

There are two explanations, and four months cannot tell them apart. Either we were unlucky with the market (ETH rose 34% in four months), or the 34.8% ceiling is lower than the real one. The number itself is also waiting for a recalculation at the current breaker threshold, more on that below. There is no comfort in this, and no panic either.

One thing: the circuit breaker threshold, now 26 days. In Letter No. 1 we warned that the threshold was being revised on a grid with real funding. Here is the result.

On the window from November 2021, where funding is known, the previous settings against the current ones:

Fewer cuts, a higher result and a smaller drawdown all at once. Results like that usually worry us more than they please us, and we were worried. We checked whether 26 sits on a ridge or on a needle: its neighbours on the grid (23 and 24, 27 and 28 days) are clearly worse, so it is high ground, not a random cell. Together with the threshold we nudged the sensitivity of the trend flag itself; we still do not publish it.

What we cannot say. This is again an optimum found on history, and again from a grid. Everything we said about searching in Letter No. 1 applies here too. We have not recomputed the Hansen test (SPA) or the Deflated Sharpe at the 26-day threshold, so the word "verified" does not fit it. The numbers in Part I of Letter No. 1 are also waiting for a recalculation, because they were computed at the previous threshold: the 1.17× / ×5.29 / ×1.48 decomposition, p = 0.028, 0.971, the 62% budget optimum and the 34.8% drawdown. The direction of the conclusion does not change (all the value and all the risk sit in the circuit breaker); the specific numbers will.

Everything else is unchanged: the 62% budget, the 3.5% target, one level, entry on a schedule, no averaging down, no pattern.

This section is not decoration. It shows that searching is paid for, not quietly accumulated.

A price stop-loss, again, on the full grid. It is the first question we get. The tighter the stop, the worse both the result and the drawdown at once: with no stop ×3.76 at a −63.2% drawdown, with a 5% stop ×0.22 and −87.6%. The explanation is simple: our breaker cuts only a confirmed reversal, while a price stop cuts any ordinary noise inside a reversal that would have come back to target on its own. It turns future wins into booked losses.

Eight ways to work inside a rally. Entering small, catching the slope turning, selling the stack on a signal, riding the price instead of a flat target, and four more. None passed. Typical numbers: a minimal 0.5% target gives 0.14× and the cuts do not go away; a fast sell signal showed +6.0%, but on the placebo test p = 0.130. One cause covers all of them: a lagging signal is good for forbidding and no good for acting. A late ban is harmless; a late action costs money in both directions. So in a rally we do not work, we wait it out, and today's account shows what the waiting costs.

Cutting less often. We looked at whether a different exit target would help. The number of cuts is set by the market, not by our target: 34 to 36 over nine years across the whole range of targets. A 4.5% target is better on history but does not carry over.

The same short with no breaker and no gate. This is no longer our account but an experiment in reverse: buy $200 a day, short $200 a day, and close everything with one order. The result is worse than simply buying (×0.86) on the full history and barely neutral since 2021 (×1.02; funding not included). The cause was found and measured: one book left open since 2020 carries about −209 ETH of paper loss against +31 ETH of realised profit from 201 closed cycles. One stuck leg is more than six times all the profit from the closed ones. That is what our account looks like with the breaker removed.

The calculation of the account's history was inaccurate, and we found it on 30 August. Two defects. First: when the history was recomputed, the last day ran with the circuit breaker switched off. An unknown trend slope was read as "no trend", entry was allowed and a cut was forbidden. Because of this, in the recomputation from 13 August a short was opened at a moment when the filter was required to forbid it. We computed the counterfactual: with the filter working, entry would have come a little later and the loss would have been 0.3 ETH smaller, so the bug cost us little, and in our favour, and the loss came from the market. Second: the prices for the recomputation were taken from the linear contract, while we trade the inverse one, so decisions were being made on someone else's candles. Both were fixed on 30 August, tests were written for each that failed on the old code, and the history was recomputed. Every number in this letter was computed after the fix.

The "worst move against a position" field is understated, and we found it while preparing this letter. When the engine restores a position from its record, it does not remember the position's worst point and starts counting again. For the trade of 9 August the record says 31.0% instead of the real 33.9%, and for the open position 3.0% instead of 15.0%. In this letter the numbers are recomputed from candles. The defect itself is not fixed as of the date of this letter.

The history cannot be rebuilt from scratch from public data. The exchange gives funding history for about three months, so a full recomputation loses funding for June and early July, and the sizes of the first positions come out 5 to 9 contracts different. The numbers in this letter come from recorded trades and recorded funding, not from a recomputation. We consider this a weak point for verifiability.

Everything said in Letter No. 1 still stands, and three points from it have just come true: we lagged in a rally, the drawdown turned out deeper, and the win rate means nothing. Let us add what was not there.

We simulated three years ahead, and the result is unpleasant. A block bootstrap over the full history (500 paths, 13 blocks of 90 days, current settings): a median of +12% a year in coin, a 5th percentile of 0.49× the stack, a 95th of 4.78×. In 30% of paths we finish with less ETH than simply holding with no trading at all would have given. Median drawdown −44%, worst path of 500 -81%. The stack was never wiped out in any path. These numbers are in ETH, not dollars: they answer "if the capital is already meant to be held in ETH, what does our structure add on top of holding", and they do not answer "crypto or fiat".

Mandatory reminders.

• In dollars, over the long run, the drawdown is of the order of -85%.
• The bias ratio in coin is 4.78 against a norm of 1.0 to 1.5. A standard screen for smoothed quotes will flag our coin series red. This is an artefact of the denomination: in dollars the figure is 1.01, and a dedicated smoothing test confirmed it. We say so before you ask.
• The return comes from the falls: on rising days we beat the market 27% of the time, on falling days 72%.
• The worst lag behind simple holding for our class of strategy is -48%, and the longest losing streak against holding is 14.6 years. Today's −9.8% in coin is not yet on that scale, but it points the same way.

And 122 days is still not a sample. Ten trades, two cuts, one open position. They show that the mechanism works as designed, and that its price in a strong rally is higher than we described. They say nothing about the next quarter. Please do not multiply −9.8% by three, just as we asked you not to multiply +5.25% by five.

The account stands at 45.10 ETH instead of 50 (−9.8%), and ETH rose 34%. The circuit breaker fired for the second time, correctly both times and expensively both times: −14.9 ETH against +9.6 ETH from the eight wins. The drawdown reached 31.0%. We recomputed the breaker threshold and it is now 26 days, an improvement on history that has not been put through the search test. We found and named our own flaws. On 1 October a third position is open and down 2.5 ETH, and once again we do not know how it ends.

In the next letter we will again not show you a dazzling new idea. We will show the same one, checked once more, with a list of what we rejected. If that sounds dull, we are working towards the same impression.

Thank you for your confidence. We are glad to answer any questions in person, including the awkward ones. There will probably be more of them than usual this time.

Yours faithfully, Isochrons Capital

In two cuts the circuit breaker cost us 14.9 ETH and saved us an unknown amount. We measure the first to five decimal places and cannot measure the second at all. This is still the main difficulty of explaining our work.

"In a rally we will look foolish" was written as a caveat. The market read it as an order and carried it out with rare precision.

We are better at finding our mistakes than at making them. So far. This time Part IV has several, and we found them all ourselves.

Past performance does not guarantee future results. Returns are calculated in the contract currency (ETH); the dollar-denominated result for the period differs and is stated separately. Single-period figures are not suitable for extrapolation. The three-year simulation rests on the history of one asset and is not a forecast.