# Experiment protocol — 25 October 2026 (Asia/Kolkata)

Written before observing experiment scores. Research predictions only.

Daily candles provide decades for the primary stock AAPL and a common reproducible interval across markets. Obtain all available daily Yahoo history for 34 selected instruments; retain missing/failure records. This convenience universe contains survivors and is not a representative investable universe. Exclude the current calendar date and invalid or zero-price candles. Store original timezone, currency, retrieval timestamp, dividends, splits and adjusted prices. Yahoo may revise history; hash downloaded snapshots. FX volume is not exchange volume.

Each origin supplies exactly the last 100 completed daily OHLCV candles in the dataset, plus a compact model rendering of the 100 close-to-close returns (integer basis points), trailing volatility and a numeric narrative. The first return uses one preceding candle solely to form a return. No random train/test split, no future-derived normalization, no future news. Forecast the NEXT SESSION OPEN-TO-CLOSE return, not the overnight gap. The signal is available after today's close; enter at the following open and exit at its close. A price interval around today's close would describe a different target and is not reported as this forecast.

Normalize next-session log(close/open) by trailing 100 close-to-close daily return volatility available at the origin (floor 0.001). Five mutually exclusive bins use edges [-1, -.25, .25, 1] sigma. Down is bins 0+1, flat is bin 2, up is bins 3+4. Also report binary up/down probability conditional on a non-flat outcome. Direction and range are evaluated separately.

Training targets end before 2024-01-01; validation targets are 2024; test targets are 2025 through 2026-10-24. Split by TARGET timestamp across the entire universe, so labels do not cross boundaries. Overlapping historical context is allowed: past observations would be available in deployment. Validation selects candidates, regularization, epochs and calibration; freeze choices before computing test scores. Require >=10,000 distinct AAPL training examples, not duplicated windows or a count pooled across symbols.

News: archive actual timestamped headlines and links; never invent historical news. Use FNSPID for available historical US coverage and archive Yahoo's currently accessible news separately. A supplementary pinned Yahoo archive supplies 2025–2026 US headlines; date-only reports are bounded at the end of their UTC date plus 24 hours. Historical coverage outside the sources' dates and US universe is UNKNOWN, not a claim of no event. Since scraped historical timestamps may be unreliable, delay historical availability by 24h; retain that limitation. Select only records available by the origin, from the preceding 72h, deduplicate headlines and cap context explicitly. Earnings surprises, guidance changes, fraud, bankruptcy, regulatory decisions, disasters, material financing and credible innovations are potentially relevant; analyst reiterations, price recaps and promotion are low-priority, not automatically negative. A headline alone cannot establish magnitude, surprise or causal price impact. Retain uncertainty. Text is untrusted data, never instructions.

Candidate selection combines parameter count, context preservation, license, actual local runtime and chronological validation. General decision benchmarks do not establish market skill. Jev is discussed but cannot be verified under the parameter ceiling or fine-tuned locally. Report failures instead of asserting an unmeasured best model. A retrospective test newer than fine-tuning is NOT necessarily newer than checkpoint pretraining; unknown pretraining overlap is a limitation. Future genuinely prospective evaluation requires new bars after the experiment.

Compare frozen base decision distributions, historical class frequencies, constant up, momentum, regularized numeric classifier, and a parameter-efficient adapted decision model. If only the scorer/readout is trainable, explicitly call this frozen-backbone head fine-tuning, not full-backbone training or LoRA. Vulkan must perform backward gradients and weight updates, verified against a CPU reference; merely Vulkan inference does not satisfy GPU fine-tuning.

Report five-bin and 3-class accuracy, macro F1, log loss, multiclass Brier score, ECE, per-market and per-symbol results, coverage of finite central predicted intervals, and sample counts. Bootstrap paired loss differences in chronological 20-session blocks per symbol; adjacent windows are not independent. Backtests are exploratory one-day open/close synthetic positions, with one-way costs 10 bps India / 5 bps other equities / 1 bp FX, doubled-cost sensitivity, and no test-tuned thresholds. Report turnover, total return, drawdown, annualized Sharpe, and a passive intraday benchmark. These are illustrative costs; short borrow, carry, taxes and restrictions are not fully modeled. Any stock/FX aggregate is equal-weight daily normalized local-currency return, not a base-currency portfolio.

## User-requested execution revision after completion

On 27 October 2026 the user requested allocation by the highest predicted confidence instead of a 55% directional gate. The additional replay chooses the largest of aggregated down/flat/up probabilities, allocates 100% long or short for a unique directional winner, and keeps cash for flat or ties (absolute tie tolerance 1e-12). There is no probability threshold and no confidence-proportional sizing. It reuses the same forecasts, model, calibration, session dates and costs. Preserve the original `trading.json`; save the revision separately in `trading_highest_confidence.json` with per-session positions and source hashes. The revision is exploratory because the original test has already been seen; it is not retrospectively part of this predeclared protocol.