NilsTrades

Testing what traders argue about.
Data, not feelings.

[NilsTrades @nilstrades]

https://t.co/EVRluQ7NMT

[NilsTrades @nilstrades]

@dmartin_trading One property worth knowing about the fixed 90th percentile boundary: With a check after every trade, any healthy strategy is almost guaranteed to breach it eventually, so the effective false positive rate is far higher than the 10% the percentile suggests.

[NilsTrades @nilstrades]

I have 75 assets in my watchlist that I theoretically want to trade, grouped into 34 correlation clusters.

I exhaust daily timeframe first as it is the best (most robust) performing from what I saw.

Each cluster has 4 "strategy slots"
- trend onset
- trend continuation
- mean reversion counter-trend
- mean reversion pro-trend

For assets that hav a long term bias (US Indicies for example), those can be split further into bull and bear.

So theoretically, there are 120+ strategies to deploy (1 per cluster per style) before reaching any ceiling, on daily timeframe only.

Once that is done, I will maybe look into intraday approaches. Or just become a full time VPS monitoring person 😂

[NilsTrades @nilstrades]

@kieran__duff The hardest right thing to do. 💔 Sending hugs.

[NilsTrades @nilstrades]

@kieran__duff Margin of error based on sample size, calculated from a 50% win rate:

100 trades = ±9.6%
(that means the actual win rate could be anywhere from 40.4% to 59.6%)
500 trades = ±4.4%
1000 trades = ±3.1%
10000 trades = ±1%

[NilsTrades @nilstrades]

@kieran__duff When forex traders boast about their 5 minute timeframe backtests. That timeframe easily turns a highly profitable strategy into a losing one, all thanks to trading costs.

[NilsTrades @nilstrades]

@QuantifiedStrat Also interested in the rules 🙏

[NilsTrades @nilstrades]

@TheH0n3stTrader Tell me you don't understand maths and statistics without telling me you don't understand maths and statistics.

[NilsTrades @nilstrades]

Your sample already contains the regime this component is meant for. 2022 was a ten-month NQ downtrend, which a trend short strategy should capture cleanly. (Fast V-crashes like 2020 do not count in this case, since a 200-bar breakout enters late and gets whipsawed.)

The debate can easily be settled with data. If removing the bear hedge reduces max dd, your friend is right.

Insurance that never pays a claim is just a cost.

[NilsTrades @nilstrades]

One thing I learned from my previous career in marketing statistics: If you look at a metric, and it doesn't give you a clear answer to "so what!?", it is costing you.

Either in time invested looking at it, or in making adjustments based on it that should not be done, trying to fix a problem thst doesn't exist.

[NilsTrades @nilstrades]

@kieran__duff Very well written. I haven't had the "pleasure" of a public track record, but this text makes me reconsider. It does sound like something to grow by.

[NilsTrades @nilstrades]

@dmartin_trading The big difference is that emotions don't make you skip the next trade, or exit the current one early. So while emotions exist (we are humans, after all), they don't get to make the choices. And I think that is a pretty huge advantage.

[NilsTrades @nilstrades]

@dmartin_trading Yes, but the posts are still visible. It's just a matter of regulation and politics and lobbyism for now. Eventually models like Fable, and models even more capable will be available.

[NilsTrades @nilstrades]

@TradeQuantiX @pedma7 Curve fitting of the future equity curve. 🤌

[NilsTrades @nilstrades]

@dmartin_trading It's funny how automated trading shifts to server monitoring. This is your future now. Embrace it. 🤣

[NilsTrades @nilstrades]

If all 10 are optimized together, that is true; that's almost always curve-fit.

It changes when they are added one at a time: Lock the primary trigger first, then add filters one by one, and lock a filter only if it improves results across assets on its own.

Each input earns its place independently, so the real degrees of freedom stay low even at 10 inputs.

Every time I optimized everything at once I produced nothing but curve fits. Adding things one by one is the only thing that gave me good results.

[NilsTrades @nilstrades]

Pressure-test your signal, not your luck.

74 markets over 20 years of daily data, merged into 35 independent clusters using Spearman correlation with hierarchical clustering.

Run it across them and count how many it survives: https://t.co/20Cf0UyeB1

[NilsTrades @nilstrades]

https://t.co/EVRluQ7NMT

[NilsTrades @nilstrades]

One layer on the retire case: A streak that breaks past the range your backtest showed is not automatically proof the edge is gone. On a small live sample a single bad run can exceed the historical worst and still be variance, because that range was estimated from finite data too. The breach only becomes a retire signal once there are enough trades to trust it. I personally investigate when 10 trades in a row have stayed below the 5th %ile of the MC simulation..