Do the signals help predict returns?

Previously studied history · no trading authorization

Can model scores identify the stocks that rise later?

The models do not yet rank later returns consistently. Five-session scores changed direction between windows, and twenty-session evidence remains limited.

Missing results stay blank. Every cost point is a separate saved account calculation. The table retains unavailable reasons.

Each model predicts the next session's first trade to the fifth or twentieth session's last trade. IC measures agreement between the score ranking and later return ranking. Zero shows no ranking relationship. MSE measures squared prediction error; smaller values mean smaller errors within the same horizon.

Training and validation use earlier dates with complete labels and strict time separation. Two first-window twenty-session models had only 14 valid inner-validation dates, below the preset minimum of 20, so neither was refitted. A later Elastic Net forecast was constant and has no ranking IC.

Scores assume end-of-session signal formation. Actual source availability, complete security types and corporate actions remain unverified.

What has been checked in the stock inputs?

Candidate-days with unresolved identity

657,731

Issuer announcements / filings reviewed

7

Raw-price boundaries matching split seeds

4

Large adjacent observed raw-price jumps

243

The input retains 658,472 candidate-days. Seven known action boundaries are quarantined without changing prices. Four split boundaries match the sparse seeds. The 243 large adjacent observed jumps include gaps and other effects; they are not a count of confirmed corporate actions. Complete type/action history is still missing.

Earlier observation preparation

The earlier free daily-archive extension stopped at its fixed 300-second limit after 345 attempts and 55,656 recorded response bytes. The receipt has uncertain final transfer spend. At that snapshot, warmup was incomplete, no new registry was frozen, and no future observations were accepted. There was no automatic retry or schedule.

Previously studied history

Stock selection after modeled costs

All completed daily stock-selection proxies lost money at 20/40 bps one-way costs. At the original timing, Ridge trailed simple reversal in both comparable windows.

The counts below cover every requested window at the selected cost and timing. Ridge exceeded equal weight in 2 of 2 comparable windows. Another 1 requested windows could not be compared.

Four methods use the same candidate data and cost assumptions. These are historical price-and-cost estimates. Security identity and executable prices still need acceptance.

20 bps means 0.2% of traded value. These are assumed total costs. Stock trades use the next exchange session. Delaying adds one more exact session. All positions close within that session, with costs on both entry and exit.

Equal weight uses all candidates with a usable signal-day price. The three ranking methods choose 50. Their concentration differs, so return differences include that risk difference.

Each window starts at normalized NAV1, denominated in USD, and requests 20 signal dates. Entry and exit use the first and last usable regular-session trade. These prices do not establish executable fills.

Does the result depend on a few opportunities?

These are ex-post diagnostics. The five highest daily returns are omitted mechanically; holdings and costs are not recalculated. The asset share uses positive price PnL as its denominator.

What the input checks establish

Daily candidate rows

658,472

Rows without usable OHLC, including missing prices

441

Candidate-days with unresolved identity

657,731

Minute sample rows

120,017

Regular-session sample rows

49,530

The minute check covers one fixed row group, including pre/post-market records. Full conversion completion is the converter's reported status.

The future observation proposal is frozen and awaiting fresh data with verified availability. Historical replays remain a separate evidence stage.

Does the model add anything beyond a simple reversal?

The extra features added little to Ridge in this sample. The shallow tree helped in the middle period and fell behind in the last one.

45 fits, up to 500 source candidates per day, three periods of 20 trading days. These dates were already studied. Historical security identities and trade execution remain unverified.

Reversal ranks today's bigger decliners higher for the next trading day. Rank IC measures how well the ordering matches the next day's intraday move. A higher value means better ranking. It does not tell us how much a portfolio earned.

Where did the ranking work?

Groups use each feature date's eligible pool. Small or constant groups remain unavailable. These comparisons describe this sample and do not identify a cause.

Daily prices do not prove order fills, market depth or trading capacity. No brokerage account was connected. The existing observation registry is unchanged.

Equity prediction study

Can a simple model rank tomorrow’s moves?

Ridge had little ranking effect in the middle test period and a stronger positive effect in the last. The three periods do not establish a dependable trading advantage.

Monthly pools contain at most 500 source candidates selected using the previous 20 exchange sessions. Ridge and Elastic Net predict the next session’s observed first-to-last trade return using six daily features.

Security identity, historical security types, corporate actions and execution remain incompletely verified. This historical period was already explored. These prediction scores carry no accepted trading PnL.

2025-09-25 ～ 2026-10-20 · 2518 source identity segments · 2511 unresolved segments · 42 fit attempts

Each test period requests 20 exchange dates. A day needs at least 30 distinct candidates and nonconstant scores and labels for IC. Zero forecasts have undefined IC. Candidate pools change monthly, so distinct identities across a period can exceed 500.

Strategy results are available; factor validation is still pending

Momentum and moving-average rules have been backtested. That does not yet show whether the underlying signals predict later returns consistently.

The stock prediction study above measures ranking separately. The four signals below still need their own IC, grouped-return and signal-decay tests.

What would establish stronger evidence?

IC / rank-IC

Check whether signal values agree with subsequent returns, using only information available at the decision time.

Grouped returns

Sort assets by the signal and compare later returns across groups, including costs.

Stability and decay

Check different periods, markets and holding horizons on fresh data.

The four signals in this section have no published IC, grouped-return or signal-decay scores. Their future tests remain separate from the stock prediction study above.

Strategy results are available; factor validation is still pending

Momentum and moving-average rules have been backtested. That does not yet show whether the underlying signals predict later returns consistently.

The stock prediction study above measures ranking separately. The four signals below still need their own IC, grouped-return and signal-decay tests.

What would establish stronger evidence?

IC / rank-IC

Check whether signal values agree with subsequent returns, using only information available at the decision time.

Grouped returns

Sort assets by the signal and compare later returns across groups, including costs.

Stability and decay

Check different periods, markets and holding horizons on fresh data.

The four signals in this section have no published IC, grouped-return or signal-decay scores. Their future tests remain separate from the stock prediction study above.

Hourly crypto factors: no real results yet

The first authenticated check returned HTTP 403. Acquisition stopped before any market data was downloaded.

The implementation passed synthetic tests. Provider access remains unverified.

Permission to publish derived research statistics has not been verified. No measured factor chart is published.

The offline tool checks four fixed signals against 6-hour and 24-hour forward returns. It reports rank correlation and descriptive groups, with sample counts. It does not simulate a traded portfolio.

• momentum_24h Price change over the previous 24 complete hours.
• momentum_168h Price change over the previous 168 complete hours.
• volume_ratio_24h_7d Volume in the last 24 hours divided by average daily volume in the preceding seven days.
• momentum_volume_24h Keep positive 24-hour momentum when the volume ratio exceeds one, otherwise zero. Missing inputs remain missing.

Powered by DexPaprika