Dogtor🎰📈

🧵 As a penny-stock swing trader, obsessed with data. I spent months building something most retail never bothers with - an 816-event reverse-split dataset, hand-labelled for a statistical deep dive to find the edge here Thesis - you can go long after the reverse split to harvest the mega pumps (>1,000%) like we saw in $TNON, $AREB, $SMX and $AKAN lately. 📁 My key intellectual asset is the .xls dataset 816 US R/S events, Sept 2024 → May 2026. Manually curated from TradingView , SEC filings, dilution trackers. The brutal part wasn't the dates - it was nailing the exact shares outstanding immediately after each split back in the day - took months, but I made it. Companies not always broadcast that number in their pre-split PR(especialy if ADS are involved). You back it out from the next 10-Q, an 8-K Item 3.02, or float-compression math inside a 424B. Sometimes three filings to confirm one row. What do I concider as a Pump⛽️ = any 1-day intraday ≥50% move within D+1–D+10 of effective date (incl. pre/post-market which is the most common). 🚨FIRST FINDINGS: post R/S pump rate - 30.3%. Sub-1.4M S/O cohort it is even bigger - 38.2%. Now I need to find additional predictors and develop a backtested trading strategy!🚨 🧪 Methodology 15 pre-registered hypotheses across float, split ratio, geography, toxic underwriters, structured-finance funds, cash runway, prior R/S, sector, VIX, entry-day etc.. Two layers: P(pump), and P(extreme | pump) - the 300%+ TNON/SMX archetype. 🔬 The pipeline Contingency tables + χ²/Mann-Whitney univariate, ORs with 95% Wald CIs, Benjamini-Hochberg FDR at q=0.10 across the family. Phase 4 (out of 10) will cover logistic regression with stability selection (500 bootstrap resamples, retention ≥60%), XGBoost under nested time-based 5-fold CV - outer loop for performance, inner Optuna for hyperparams (no tune-on-test leakage), SHAP for direction, RuleFit for lift-≥2 / support-≥30 rules, decision-curve analysis for net trader gain, reliability + Brier on a time-held-out test set. 📊 So far I'm at Phase 2a and preliminaty result are really encouraging. 🚀 What's next? Hypothesis 1 held! Already enriching - toxic-fund 13Gs, last-424B underwriter ID, going-concern flags, warrant overhangs (edgar data collection). If the multivariate gate (PR-AUC ≥0.30, top-decile precision ≥30%, out-of-sample) holds, Claude builds a Claude Skill and a Dashboard so I can automate the future research: ticker + R/S date (usually given in advance) → instant pump probability with SHAP reasoning 🔒 Disclosure: I made sure, the research is not replicable from this post as the signal lives in the locked .XLS dataset (only dataclearance took me one week🤓) and you can see the stevendu acknpwledgment of the key constrain for such project. If Phase 4 fails the gate, I close with a null-result write-up. That's still on the table. Let me know below if you are big on coding / statistics / have tons of Claude tokens😄 to participate in this research and as a result get access to the destilled trading strategy and finally find your statistically-proven edge, gave up gambling 🫂