Research

A news classifier can be 96% accurate and miss most relevant stories

Our GDELT probe reached 95.90% overall ticker-array agreement while matching only 7 of 20 relevant headlines. The denominator explains the difference.

Open page

AI trading agent safeguards: stop the next trade before a bug compounds

Watch an order wait for approval, trace a buy/sell loop through its events, and see a daily allowance stop further automatic trades in NexusTrade.

Open page

Are news headlines enough for stock sentiment analysis?

Our broad GDELT audit exposed missed parent companies, conflicting model scores and faulty labels. Here is what article text might fix and what needs a separate repair.

Open page

Backtest data sources and pricing checks

Follow a recorded SPY option decision and fill, inspect Massive’s documented exchange feeds, and check quote timing and historical holdings availability.

Open page

Best and worst trading indicators in the offline backtest corpus

Compare technical indicators by median Sharpe in the available September NexusTrade archive, with sample counts, outcome dispersion, drawdown and reproducible cohort rules.

Open page

Can AI Pick Stocks? 12 Years of AI Stock Ratings vs. Returns

An AI model grades thousands of US stocks from 1 to 5 on their financials. See what the score measures, what it misses, and how each rating group's stocks returned over 12 years.

Open page

Can an LLM Edit Neural Network Weights? My Qwen Grader Experiment

Reproduce the published Qwen weight edit, inspect a nine-case agent-grader experiment, and compare saved edits with the original model and random controls.

Open page

Can news sentiment improve a top 10 stocks strategy?

I filtered the 10 largest US companies by 60 sessions of news sentiment. The worst drawdown fell from 41.4% to 31.6%, yet holding the top 20 with no news did better, and news on the top 20 cut its return.

Open page

Check a public ownership finding against its filing

Follow one Apple holding back to Berkshire’s filing, check its dates and units, then preserve a research note another reader can verify.

Open page

Compare stocks and discover similar companies

Find businesses with shared core activities, then compare their price charts, valuation and financials. Similarity scores come from annual filings released in 2025.

Open page

Does copying Congress beat SPY, and does entry timing matter?

We tested congressional stock purchases against SPY and compared entry before and after disclosure. Explore both findings, the uncertainty, and results by politician and disclosure delay.

Open page

Does LLM Hindsight Change News Sentiment Scores?

I scored 10 famous stock headlines and summaries with Claude Haiku 5.5, as written and with the company masked. The mean masking gap (0.15) beat mean rerun noise (0.10), and masking also deleted information, so the result is inconclusive.

Open page

Does news sentiment predict stock returns? Results across five horizons

We tested 2.27 million company days across the next session, week, month, quarter and year. See the return gaps, uncertainty, missing outcomes and the next hypothesis.

Open page

Five dividend growers: rules and three completed backtests

Inspect an existing five-stock dividend strategy, its original rules, costs and three completed historical periods, including its longer-period loss relative to SPY.

Open page

Genetic optimization: follow a real NVDA search

See the starting NVDA rules, actual genetic changes, later backtests and filled order. Open the complete experiment in Aurora.

Open page

How Do Congressional Stock Trade Disclosures Work?

How NexusTrade turns House and Senate trade reports into dated rows, the PoliticalTrades indicator and copy-trading bots, and how often each updates.

Open page

Lookahead bias: audit signals, filings and historical fills

Follow one decision made before its input existed. Compare the early and corrected timelines, then ask Aurora to check the exact timestamps.

Open page

Mean reversion vs momentum: what the existing corpus can compare

Compare buying a drop with buying strength. Follow two explicit teaching rules, then inspect what the historical archive can and cannot compare.

Open page

Out-of-sample testing: inspect a frozen NVDA strategy

Completed NVDA experiment: a frozen searched strategy gained 22.06% on later dates, held overnight and failed four of its five daytrading requirements.

Open page

Point-in-time universes: dated membership and dividends

Audit historical stock eligibility with Tesla's dated S&P 500 addition, Target dividend announcements, and a reproducible worksheet for the five-dividend-growers study.

Open page

Portfolio risk indicators: read the path behind the return

Follow a $3,500 holding inside a $10,000 account. See the concentration, check the proposed cap, and ask Aurora to explain the exact sizing before changing a rule.

Open page

Review one public ownership disclosure

Start with Berkshire’s reported Apple position. Learn what to compare before calling a holding new, increased or sold, and open the source in NexusTrade.

Open page

Selection bias in strategy research: keep the losers in the record

The same dividend strategy beats its stored SPY baseline in one period and trails it in another. Follow both actual curves before choosing which result to share.

Open page

Stock news sentiment analysis: fine-tuning a cross-encoder for trading

Fine-tuned DeBERTa agreed with reference labels on 87.9% of 434 company validation examples, versus 74.5% for frozen Qwen embeddings. I scored 67.8M target pairs; trading results are next.

Open page

Technical vs fundamental analysis: work through one Google decision

Read a dated Alphabet earnings example, inspect actual Google strategy rules and separate a business thesis from a price-based entry condition.

Open page

Trading condition rankings from the existing backtest corpus

Compare complete condition trees, operand assets, thresholds and allocation actions within dated ticker and execution cohorts from NexusTrade's existing offline backtest corpus.

Open page

Trading strategy overfitting: inspect a real earlier and later test

Compare the same NVDA strategy on earlier and later dates, then inspect the changed rules and actual orders before explaining the failure.

Open page

Trading strategy rankings from the existing backtest corpus

Compare complete stored strategy cohorts from NexusTrade's existing offline corpus by median Sharpe, Sortino, return, drawdown and paired benchmark excess return.

Open page

XGBoost vs neural networks for sentiment analysis: which is better?

I compared XGBoost with a small neural network on the same financial headlines and embeddings. The MLP had 10.4% lower sentiment error at the largest training size.

Open page

Continue exploring

Browse the complete library ↗