Joined December 2023
- Tweets 1,596
- Following 0
- Followers 424
- Likes 5
418 Photos and videos
Another fund raising seven figures to throw foundation models at markets. But financial time series are mostly noise, not language with rich structure. Bigger models overfit faster. The bottleneck was never compute. It is a signal that survives costs and capacity.
48
'React, don't anticipate' sounds wise but it's untestable as stated. Make it a rule: enter on confirmation, not forecast, then backtest both and compare Sharpe net of costs. If feeling the market's flow beats a coded signal, prove it. Usually it's just hindsight talking.
44
Oldies but Goodies
Machine Learning Risk Models build equity covariances from averaged k-means clusterings of recent returns, not a factor model, giving an invertible matrix for portfolio optimization.
paperswithbacktest.com/strat…
53
Backtests treat every fill as real and final. But exchanges keep an "obvious error" rule: clearly erroneous trades get busted hours later. Your best print of the day can simply vanish. Edge that only lives in fills no human would honor was never edge.
1
70
The biggest tail risk in a trading system isn't a bad signal. It's wiring untested, high-blast-radius tooling into your live execution path. Operational risk earns no Sharpe, but its drawdown is uncapped. It isn't backtestable, which is exactly why people skip it.
60
A 7-year streak of positive annual returns feels like proof of edge. It usually isn't. Shift the start date and most low-vol strategies show one. The testable question is never the streak. It's Sharpe net of costs, worst drawdown, and out-of-sample stability.
47
Oldies but Goodies
Build a global expected business condition factor from lagged OECD leading indicators, then time international equity markets monthly. It predicts aggregate returns in and out of sample, working through the cash-flow news channel.
paperswithbacktest.com/strat…
36
A stretched pairs spread looks like free money. Before you fade it, ask: did the relationship break, or just deviate? Most stat arb mispricings are regime changes wearing a mean-reversion costume. The edge isn't spotting the gap, it's knowing it closes.
1
1
67
Bitcoin trades 24/7 — and that never-sleeping market leaks a pattern. From Oct 2015 to Feb 2022, simply buying BTC at 21:00 UTC and selling at 23:00 UTC each day returned ~33% annualized. Two hours a day. 🧵
1
65
The insight that ties it together: a new high and a new low can't happen at the same time. So the momentum and reversion signals are mutually exclusive — you can fuse them into one strategy. The result: the smoothest equity curve of all.
1
33
Full breakdown Python code and backtest results → blog.paperswithbacktest.com/…
Explore all strategies and datasets on paperswithbacktest.com
38
Every quant believes the data hides an undiscovered edge. It mostly hides noise. The few real signals are already crowded and decaying before you arrive. Finding alpha was never the hard part. Surviving costs, capacity, and decay is.
1
2
82
Spotting two assets move together on a single day is a Rorschach test, not a signal. Cross-asset correlations flip sign constantly across regimes. The only question worth asking: does conditioning on oil actually lift the Sharpe of a small-cap timing rule, net of costs?
48
Oldies but Goodies
When wages are stickier than productivity, high labor-share firms have shock-sensitive profits and higher returns. Long the top quintile, short the bottom.
paperswithbacktest.com/strat…
36
Using Claude Code or Cursor to build a trading system automates the easy part: writing code. The hard part is unchanged. An edge that survives costs and slippage, and the discipline not to overfit. Faster tooling just produces overfit backtests faster.
2
3
68
Most option-selling backtests die from margin, not drawdowns. Your Sharpe assumes constant exposure; SPAN doesn't. When vol spikes, the broker forces liquidation at the worst moment. If your backtest doesn't model the margin curve, the strategy doesn't exist.
1
1
71