The spot side of the account bought on a dip rule: hold the incoming cash, then spend a tranche of it on a day when the thirty day trend was negative and the last two closes were each below the one before. We wrote that rule ourselves and we liked it. Spending into weakness felt like the careful choice.
On 21 August we measured it against the dullest alternative we could think of: spend the inflow the day it arrives. Daily ether closes since 2015, the dip rule read out of the running code rather than tuned for the test, five policies in all, two of them hybrids that split the inflow between the two ideas.
The dip rule did pay a lower average price. It also bought far less coin. Over the exchange traded fund era it finished with 41.8 per cent fewer coins than spending daily. In the bull window from October 2020 it finished with 86.4 per cent fewer, and at an average price 58 per cent higher than the daily policy paid. Waiting for a dip in a rising market is buying later and higher.
The better price then turned out to be mostly an artefact of how we were measuring. Set the spend equal and the advantage nearly disappears: the hybrid that sends half the inflow out every day spends 2 per cent less than the daily policy and pays 0.5 per cent less per coin. That is the timing edge, all of it. The rest of the gap was there because the dip rule never spent 45 per cent of the money at all, and an average price computed over money spent rewards a policy for not participating.
The reason is one line of arithmetic that we could have written before building anything. Throughput = probability of a dip × tranche size. Dips were 12 per cent of days, so the rule could absorb about half of what was arriving. The remainder sat as cash, by construction.
Cash also fails at the second job. Our short is sized off the coin we hold, because the hedge gate keeps the short at or below the spot position, so a dollar waiting for a dip raises no ceiling at all. Over the same period the dip rule would have left us carrying a position 49 per cent smaller. That is the more expensive half of the result, and the half a price comparison cannot see.
One case survived. A single sum already in hand, spread over dip days rather than spent at once: across 3,819 start dates on a 180 day horizon, the dip version bought more coin in 48 per cent of them, with a median ratio of 0.984 and a fifth percentile of 0.130. In 18 per cent of starts the money was not fully deployed within six months. A coin flip with a bad left tail is not an improvement, but it is not the loss that the flow case is.
The inflow now goes out the day it arrives. The dip mechanism stays in the code, switched off, for the one case it is good at. We kept the ability and stopped paying for it.