Investment research methods
AI models can produce forecasts, and some studies find predictive information in particular datasets. That does not make future prices certain or establish that a public tool will earn a profit. Start by defining exactly what the forecast predicts and when it could have been acted on.
The linked screener applies filters to company data. It does not predict returns or assign an AI investment score.
These are selected studies, not a systematic review, a comparison of commercial products or a Pineify performance test.
Empirical Asset Pricing via Machine Learning compares methods for estimating equity risk premiums. Trees and neural networks improve on the study’s regression benchmarks, and the paper reports economic gains in its portfolio tests. It does not directly compare a retail AI assistant with human stock pickers.
Read the Review of Financial Studies paperCan ChatGPT Forecast Stock Price Movements? studies responses to news headlines. This version distinguishes a non-tradable initial reaction from subsequent drift and reports declining strategy returns as LLM adoption rises. These distinctions matter when interpreting headline accuracy claims.
Read the specified research versionInterpretation: evidence for a method on a stated sample cannot by itself establish live, after-cost performance for a different model or tool. Read the paper’s target, information cutoff and execution assumptions before applying its findings.
A forecast of tomorrow’s closing price must be compared with a meaningful baseline, such as today’s price. A small price error can coexist with poor trading decisions, especially when daily changes are small relative to the price level.
A model may classify whether a return is positive or negative. Its accuracy depends on the class balance, horizon and evaluation sample. If the measured reaction already occurred before a trader could act, correctly classifying it is not a tradable forecast.
A cross-sectional model may rank stocks rather than predict exact prices. Evaluate the resulting portfolio, including turnover, risk exposures, short-sale feasibility and costs. A ranking is not a promise that every high-ranked stock will rise.
Require a forecast horizon, target and comparison method. Directional hit rate, forecast error, out-of-sample R-squared and portfolio return answer different questions. A 1–10 label is not automatically a calibrated probability.
Use only information available at the prediction timestamp. Financial restatements, delayed filings, future index membership and a language model’s training data can introduce leakage. State the limitations when training-data overlap cannot be ruled out.
Separate training, model selection and evaluation in time. Keep a final period untouched until the design is fixed. Report results by period and market condition rather than relying on one aggregate statistic.
Specify when an order could be sent and at which price it could fill. Include spreads, fees, slippage, financing and borrowing where applicable. A news reaction measured before execution is not available profit.
Record forecasts before outcomes are known, retain failed predictions and compare against the original baseline. Changing data coverage, company behavior and market conditions can change performance.
Suppose ten trades each risk a fixed dollar amount. Six gain $1 and four lose $2: the hit rate is 60%, but the total is a $2 loss before fees. This is arithmetic, not a backtest or a claim about any model.
A point forecast conceals uncertainty. If a system provides probabilities or intervals, check calibration on later data and the same forecast horizon. An explanation generated after the outcome cannot validate the original forecast.
The cited research does not justify “AI always wins,” and it also does not justify “no model can ever predict returns.” Treat each result as evidence about a particular task, dataset and implementation.
The free company screener sends explicit filters to Financial Modeling Prep and requests quotes, trailing-twelve-month metrics and annual analyst estimates for returned symbols. Its default request returns at most 20 matches. It does not run an AI stock-selection model or create a proprietary 1–10 predictive rating.
Separate data requests may have different timestamps or missing fields. Check the quote time and estimate period; an analyst estimate is a forecast, not a reported result. A matching company is a research candidate, not a recommendation.
Inspect company dataIt can generate forecasts, and research has found predictive information for specific tasks. Accuracy must be defined by the target, horizon, baseline and test sample. It does not establish certainty or profitable execution.
There is no model ranking established here. Compare models on the same point-in-time data, target and untouched evaluation period, including a simple baseline. Results from different datasets are not directly comparable.
No. The sizes of gains and losses, costs and exposure matter. The hypothetical ten-trade example on this page has a 60% hit rate but loses money before fees.
No. It measures forecast-error performance relative to a specified benchmark. It is not the percentage of correct trade directions, and its value depends on the target and benchmark definition.
No. It retrieves companies matching filters and requests additional financial data. Analyst estimates are third-party forecasts; they are not predictions generated by this website tool.
Educational content, not personalized investment advice. Forecasts and historical tests can be wrong, and investments can lose value. No return or forecasting accuracy is promised.
