Weekly Quant Trading Research Papers | Aug 19
The Best Algorithmic & Quantitative Trading Research Papers of the Week in 2026!
Welcome back qunats!
I created this Substack so you can easily find weekly quantitative trading research and papers in one place.
VIX and Trend Following Revisited: Nearly a Decade of Out-of-Sample Evidence
The 2017 VIX-adjusted trend model gets tested on August 2017 through June 2026, a period that did not exist when the original was written. The concentrated Top 1 version returned 14.09% net a year against 9.87% for a fixed ten month lookback, with a net Sharpe of 1.03 versus 0.81, while the diversified Top 2 version showed a net CAGR difference of negative 0.11 points and an Information Ratio of 0.01. Worth reading for the rolling charts, which show most of the Top 1 advantage came from a cluster of months in 2020 and 2021 rather than from a steady edge.
Your Research Agent Is an Undisclosed Factor Exposure, And So Is Everyone Else’s
Eight independent LLM research runs produced composite books correlated at 0.62 against a random expression baseline of 0.15, which is 1.51 effective independent books out of eight. Four published anomalies, one month reversal, one week reversal, betting against beta and idiosyncratic skewness, appeared in all eight runs. The author tests the shared weights explanation and rejects it, since the published anomaly literature control correlated even higher at 0.82.
Stop Loss vs No Stop: Which One Wins on a Mean Reversion Trading Strategy?
Adding a 5% stop to the same entry signals dropped annual return from 8.22% to 1.05% and pushed the win rate from 66.71% down to 49.05%. Max drawdown moved the wrong way too, from -18.63% to -41.78%, so the stop did not buy the protection it was there for. Twelve stop levels were tested and the full rules plus the RealTest script are in the post.
Get Papers for Quant Traders’s stories in your inbox
Join Medium for free to get updates from this writer.
AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search
The unit of search here is the whole research artifact, hypothesis, executable expression, platform evidence and review status, rather than the formula on its own. Reported WorldQuant BRAIN deployments reach a fitness of 9.50 and a Sharpe of 3.48 across several users and model backends, with an adversarial reviewer that re-executes what it is given. The useful part for testing is the traceability from prompt to final expression, not the platform score.
Self-Supervised Auxiliary Task Discovery for Stable Reinforcement Learning in Stock Trading
Auxiliary tasks in RL trading are normally hand designed and locked to a fixed target and prediction horizon, which stops fitting once the regime changes. This framework has a second network generate the task definitions itself through learned cumulants and discount factors, updated by a meta gradient on long run trading performance. Tested on DJI, FTSE, Sensex and TAIEX, and the claim is steadier learning against existing baselines rather than a higher headline return.
Sectoral Intramonth Momentum Cycle: Exploiting Turn-of-the-Month Patterns in Sector ETF Strategies
Nine Select Sector SPDR ETFs from December 1998 through June 2026 show three separate legs inside each month, trailing 252 day momentum paying on the first trading day, reversing sharply on days two and three, then a third independent leg from ten to five trading days before month end. Trading all three as one composite delivers 5.99% annualized at a 0.55 Sharpe long short, and 3.77% at 0.54 for the market neutral variant. Both stay invested fewer than half the trading days each month, so the turnover and cost assumptions matter more than usual.
1.28 Model-Based or Data-Mined: Lotter’s Framing of the Whole Problem
Every strategy gets reduced to one equation, price equals previous price plus a function of recent history plus shocks, with inefficiency defined as that function being non zero. The part that changes how you read a backtest is the detection arithmetic, a 5% momentum autocorrelation gives 0.4 pips of drift against 10 pips of noise, so roughly 2,500 bars are needed before the signal reaches twice the noise. It also covers scaling drawdowns across horizons with the square root rule, where a 12% drawdown over three years should be expected near 17% over six.
Model predicted risk from news sentiment is fed into the covariance matrix rather than only into expected returns, tested on Russell 2000 equities under three selection regimes. The strongest conservative row is pure beta selection with a 40 day holding period and risk parity allocation, reaching a Sharpe of 2.33 at 100 bps of cost. The separated pure alpha and pure beta legs usually beat the combined trigger, and the paper’s own read is that the selection regime and the allocator mattered as much as the sentiment model. :)
