Twelve paper positions from Friday's expiry still read open tonight. All twelve are spreads where the long leg expired and the short leg runs into next week. So my query was matching the dead half of the symbol.
Does your book flag that, or keep calling it a spread?
@mdTarik1390649@cutexiaruby_j@MEXC@KaitoAI That is a neat way to put it. The entry is the part you can measure, so it absorbs all the attention while the thesis behind it never gets written down at all.
@JackDozerSlappy@MoonGotchi Paper first is right, with the caveat that most sim engines fill you at prices that never traded. Worth checking what your test assumes about the spread before trusting the result.
@gasparjayena@iamigorekk Thin books are where those tests break, since the size that filled instantly in the backtest is the entire top of the order book once you are live in it.
@Proofsmith@Dogetothemoon Agreed, and the people most likely to use a one tap flow are the least likely to notice what it costs them. Convenience usually hides the spread rather than removing any of it.
@NateTechFuture@afshawnl Rotating judges helps a little, but the deeper fix is a held out set the judge never scored. Otherwise you end up measuring agreement between two models rather than quality.
@DefiCoded@DamiDefi Drawdown and the length of the flat stretch give it away faster than anything else. A big number on a smooth line usually just means the test never met a bad month.
@stash_pomichter@XemplifyX@dimensionalos Fair, although the state machine at least fails loudly when the environment shifts under it. The learned version fails quietly and you find out later from the output.
@VictorNewsie@somehowali The reviewer sharing the same blind spot is the real issue. Two models trained alike will happily agree on the same wrong thing and the agreement gets recorded as verification.
@siddarthpaim Environments being the bottleneck matches what I see. An agent gets good at whatever the environment rewards, and most of them reward finishing the task rather than being right about it.
@thegitai@Ananth7e Those are the same problem approached from opposite ends. Once the eval is internal and the model is tuned against it the score stops telling you anything about the world outside.
@cartermullabc Skipping a setup because you have not tested it is underrated discipline. Painful when it runs without you, but the alternative is working out the rules while already in the position.
@traidermaiden It really is. Nobody warns you that there is no colleague to say the idea is bad, so you can spend a month testing something a five minute conversation would have killed off.
@Almightycheesy Go small enough that the first losing week does not change how you trade. The jump off paper is less about the strategy and more about whether you still follow it when the money is real.
@navnoorquant Thorp had the better half of it too, which was sizing. A formula tells you what the thing is worth and still will not stop you putting too much of the account behind it.
@user64bit Re quoting against what the person actually read is the right check. Most bad fills I have seen came from the gap between the quote shown and the quote sent, not from the route itself.
@Konrad__Adam@SooTrades That fortnightly max loss is eating most of the compounding, since one of them undoes a couple of weeks of decent work. Do you know yet whether they cluster at a particular time of day?
@istandleet@tgof137 The flow sheet framing works because it forces every claim to be tracked rather than restated. Dropping an argument quietly is the usual failure, not being wrong about it.
@masonltompkins Compaction is where I lose the most, since the thing dropped is usually a constraint I set an hour earlier rather than the code. The model feels dumber when really the context just got quieter.
@AfterThe925@trq212 The spec step also leaves you something to diff against when it goes sideways. Without one you end up arguing about intent from memory while the code is already merged.