ragnarok

[ragnarok @ragtrader550]

@OrthogonalAlpha @therobotjames @sasuke___420 @xjpx888 @QuantumHoneybee @8020arbitrage @quant_arb @_wdearden @FormulaDeltaOne use (ln(p_offer/p_bid)) as a feature and subsample only 1 tick spreads for training data. run on >= 1 ticks.

throw ln(P_1) and ln(P_2) in as regressors and hope OLS figures out you want ln(P_1) - ln(P_2).

run PCA regression, but include all the low variance noise components.

[ragnarok @ragtrader550]

@hubermanlab what CBD/THC ratio is needed to avoid memory loss or other cognitive problems in adults, is any level of THC safe at all, and what's the level of confidence in the answer (ie how mature is the research on this question)

[ragnarok @ragtrader550]

@therobotjames i find that is often overfitting onto the holdout data which can look like alpha decay in prod

if it breaks down immediately i assume the alpha was never there unless i can observe a reason for it breaking down like a regime change in volatility or something

[ragnarok @ragtrader550]

@gwern @DegenRolf Maybe true for niche goods (fixed supply necessities) but often the market doesn't have a mechanism to equilibrate away utility surplus. Smartphones deliver $10k+ value but cost $1k. That's the subjective theory of value. Utility is usually created unless it's a demerit good.

[ragnarok @ragtrader550]

@Austen and people like them because of their personality and get angry at you when you cut them loose.

price's law + uncertainty in hiring -> hire people with large sampling variance (not just mean estimate), then either cut them or promote quickly as you get new info.

[ragnarok @ragtrader550]

@HardcoreHistory @aukulelecowboy read Jonathan Haidt, he has the best take on where the divisions come from and how to fix them.

[ragnarok @ragtrader550]

@JeromySonne @timpac @nickcammarata @Austen high up in turmeric/curcurmin supply chains, they often add lead chromate, which damages IQ. also arsenic and mercury is often found in unsafe concentrations. those herbs can cause herbal hepatotoxicity if done daily, even if not contaminated.

[ragnarok @ragtrader550]

@nickcammarata @Austen to add, intense exercise speeds up metabolism a bit.

taking up smoking does even more (halved half life).

an actual safe pharmaceutical that can speed up metabolism would definitely have market fit. i want to drink coffee late without impacting sleep quality, but can't.

[ragnarok @ragtrader550]

@nickcammarata @JeromySonne @Austen i'm considering cocoa instead of coffee during weekdays. it has a different adenosine antagonist to caffeine with a slightly longer half-life so you're not getting the crash 2 hours before the work day finishes. must detox over weekend though otherwise it builds up in the body.

[ragnarok @ragtrader550]

@nickcammarata @Austen is this also a problem for ltheanine in green tea + boiling water

[ragnarok @ragtrader550]

@FormulaDeltaOne step 4: the 10 interviewers' ratings goes into an ensemble model, and we're lifting your offer if it spikes over threshold

[ragnarok @ragtrader550]

@FormulaDeltaOne my mindset is to falsify the hypothesis as quickly as possible, unless i have good prior reasons for strongly expecting there to be alpha. sometimes a visual inspection of a chart or raw data is enough for that. sometimes a basic correlation matrix or LR.

[ragnarok @ragtrader550]

@FormulaDeltaOne because in our job we are mesa-optimizers and it's usually better search policy to do a shallow search to start with. we'll find better optima that way. no free lunch theorem and all but the search space of viable strategies is such that that's usually the best.

[ragnarok @ragtrader550]

@FormulaDeltaOne hook up GPT-3 API to twitter API

[ragnarok @ragtrader550]

@daytradingzoo Cool way to viz it, well done.

Similar thing could be done w/ non-time based buckets, too. So instead of bucketing by each month, you have volume or gap% buckets. Then recreate the viz based on those buckets to see if there's persistent non-time based variation.

[ragnarok @ragtrader550]

@CorpusCalosseum @robertskmiles @gavinmuellerphd survivorship bias?

[ragnarok @ragtrader550]

@OrthogonalAlpha makes sense, it's like a fuzzy way of firing one given the other, similar to products of hinge fn's in MARS. it's basically a regularized instantaneous pearson correlation estimate if both a,b are N(0,1).. a*b is the instantaneous correlation and exponent is the regularizer

[ragnarok @ragtrader550]

@OrthogonalAlpha so n*(n-1)/2 features, then cut down by prespecifying which pairs make logical sense. I guess we can also multiple the geometric mean by the sign of the product for directionality, and the default 0.5 exponent can be optimized for the desired kurtosis if model is additive.

[ragnarok @ragtrader550]

@OrthogonalAlpha If it involves a lot of annoying data cleaning, that's a point in its favor. If it's an unconventional idea, that's also a point. If it requires custom research infrastructure, that's a point. Filters that cut out the lazy and unoriginal from the set of competitors.

[ragnarok @ragtrader550]

@ShortDaPos In!