How Aizevin Stocks forecasts are made, measured and limited
Status: the system is under evaluation. This page explains what is forecast, how performance is measured, which records are public, and where the evidence is incomplete. Forecasts are model estimates, not guarantees. Market conditions change, losses are possible, and past or backtested results do not guarantee future returns. Nothing here is investment advice.
Quick links: Sample forecast · Track Record · Research Flow · Stock Direction
What is forecast, and over what horizon
Stock Direction model (public records). For AAPL, MSFT and NVDA, the model forecasts whether the next trading session's adjusted close will be above (UP) or below (DOWN) the signal-day adjusted close. Horizon: 1 trading session. Prices are in USD, adjusted for splits and dividends, from Yahoo Finance via yfinance.
Research Flow / Daily Dashboard (internal engine). Daily direction labels (UP / DOWN / FLAT) and predicted returns for a changing list of stocks, scored on the next session. These records come from an internal database and are not yet fully published.
Data and model
Stock Direction uses five years of daily adjusted prices (requested from 2021-09-12; latest download 2026-09-12 12:53 UTC). Each forecast record stores its ticker, signal date, signal close, model score, direction and number of training rows (for example, 912 rows for the 2026-09-11 forecasts). The model score (prob_up) is not calibrated, so it is not shown as a confidence or probability.
Missing or stale inputs are shown as unavailable rather than filled in. A full model-version registry is not yet published.
Evaluation period, sample and inclusion rules
| Item | Stock Direction evaluation |
|---|---|
| Type | Retrospective walk-forward backtest. Each row uses a model trained only on data up to its signal date (column trained_through). The results were computed afterwards, not recorded before the outcomes. |
| Dates | Signal dates 2025-09-11 to 2026-09-10; target dates up to 2026-09-11 |
| Number of forecasts | 753 (251 sessions × 3 tickers) |
| Inclusion | Every trading session and every ticker in the window. Nothing is excluded or selected. |
Metrics, baseline and results
Direction accuracy = correct direction calls ÷ evaluated forecasts. A call is correct when the predicted direction matches the sign of the next-session adjusted-close return.
Baseline: "always UP", which predicts UP every session.
| Scope | Correct / n | Accuracy | Always-UP baseline |
|---|---|---|---|
| All three tickers | 372 / 753 | 49.4% | 51.7% (389 / 753) |
| AAPL | 135 / 251 | 53.8% | 54.2% |
| MSFT | 115 / 251 | 45.8% | 50.2% |
| NVDA | 122 / 251 | 48.6% | 50.6% |
Result: over this window the model did not beat the always-UP baseline. Directional accuracy also says nothing about profitability.
Forecasts recorded before the outcome. The latest forecast file (generated 2026-09-12 13:20 UTC, before the 2026-09-14 session) holds 3 forward forecasts. That is far too few to measure accuracy. The internal production feed shown on Stock Direction has 9 evaluated forecasts (6 correct, August 2026), which is also too small a sample to draw conclusions from.
Failure cases
The weakest months in the backtest were June 2026 (26 / 63, 41.3%), April 2026 (27 / 63, 42.9%) and September 2026 (9 / 21, 42.9%). The model predicted UP in 516 of 753 cases (68.5%). Its 237 DOWN calls were right only 110 times (46.4%). MSFT was below 50% for the full year. Sharp reversals and news-driven sessions are not modelled.
Selection disclosure: Track Record and Research Flow
The Track Record and Research Flow history show selected sessions only. These are sessions from the internal results table (T_53, scope SELECTED) with direction accuracy of 60% or higher. For example, the five sessions shown from 2026-08-31 to 2026-09-08 total 31 correct out of 45 (68.9%). Sessions below the threshold are hidden, so this figure overstates typical performance. It is not used as a performance claim on this site.
P&L figures
Dashboard P&L figures are simulated, not realized trading. Assumptions: long-only positions in UP-labelled stocks, equal weight across a $100,000 notional, and a 2 basis-point cost per trade. Predicted P&L uses the model's predicted return. "Realized" P&L is implied from the recorded actual next-session price, assuming entry at the signal close and exit at the next close. Slippage beyond the 2 bps, market impact, taxes and financing are not modelled. The P&L covers selected sessions only. For example, across the 5 selected sessions the predicted P&L was $98,667.66 and the implied P&L was $7,391.80, a large gap.
Source records and reproducible calculation
- stock-direction-evaluation.csv: all 753 backtest rows (signal_date, target_date, ticker, prob_up, prediction, actual, correct, next_close_return, trained_through)
- stock-direction-forecasts.csv and stock-direction.json: latest forward forecasts, monthly accuracy, prices and source metadata
- stocks-flow.json: the selected-session dashboard payload, including provenance and the selection rule
To reproduce: accuracy = mean of the correct column (372 / 753). Baseline = share of rows where actual = UP (389 / 753). Per-ticker figures come from the same calculation grouped by ticker.
Limitations
- Only three large-cap tickers and one year are publicly evaluated. Results may not generalise.
- The backtest is retrospective. A long, unselected record of forecasts saved before their outcomes does not yet exist.
- Internal-engine results are selected and come from an internal database. They cannot be reproduced publicly yet.
- Model scores are uncalibrated. No numerical confidence is published.
- Experimental features may carry a "Beta, not for trading" notice. This page documents current practice and known gaps.