01 · Historical winner
Candidate 1
NQ-LAD
The frozen production winner. A LAD median regression using the pre-market NQ futures return.
130.2191OOS MAE · points
Explore Candidate 1 →
Quantitative forecasting research
Forecasting the Nasdaq. Before it opens.
An open forecasting experiment testing whether information available before the US market opens can improve forecasts of the Nasdaq Composite closing level.
Research and portfolio project · Not financial advice
Scroll to explore
Quantitative forecasting research
Graham Says, Model forecast for the Nasdaq Composite close:
LATEST GENUINE PROSPECTIVE NASDAQ COMPOSITE FORECAST
25,989.49
Thursday, 10 September 2026
Publication timing loading…
Actual close
26,081.72
Absolute error
92.23 pts
Percentage error
0.354%
Graham Says began as a forecasting exercise and evolved into a prospective experiment testing whether pre-market information can improve short-horizon Nasdaq Composite forecasts.
Research purpose
This is an experiment, not a trading service.
Graham Says exists solely as a research, educational and portfolio project. The forecasting output is used to demonstrate analytical modelling, reproducible research, cloud software, automated data pipelines and web development.
Research project only.
Nothing published on this website is intended as financial product advice, a trading signal, an investment recommendation, or a recommendation to buy, sell or hold any security, derivative or other financial product.
Method
One model. Frozen before the result.
M1-v1.0 is a conditional-median forecasting model selected using historical out-of-sample testing. Its specification is frozen before prospective observations are scored.
01
Observe futures
Measure the Nasdaq-100 futures return from the prior US cash-market close to 09:15 New York time.
Explore process →
02
Create a forecast
Feed the locked pre-market return through M1-v1.0 to produce one Nasdaq Composite closing-level forecast.
Explore process →
03
Score after close
Keep the forecast frozen, observe the realised close, and record the resulting forecast error.
Explore process →
PROSPECTIVE MODEL TOURNAMENT
Do the historical results hold up live?
Candidate 1 won the historical model-selection tournament and remains the frozen production specification. Separately, all five frozen model rules now compete on genuinely new market dates.
Waiting for the first jointly scored forecast.
Rankings use mean absolute Nasdaq Composite closing-level error on the common prospective sample. Lower is better. Early rankings may be highly unstable.
Model tournament
The model race.
Five pre-specified models competed on the same 1,623 rolling out-of-sample forecast dates. Scroll to move through them in finishing order. Lower mean absolute error is better.
M1-v1.0 remains the official frozen Graham Says model.
This is the historical model-selection tournament, not a retroactively constructed live competition.
The prospective race is recorded separately.
Forecast chart
Model forecast vs the market.
Hover or tap the chart to inspect values.
Prospective results
Track the experiment as it happens.
Genuine forecasts issued before the target session opens are kept separate from historical backtests and retrospective system tests.
| Date | Forecast | Actual | Absolute error | Error % |
|---|
Research · Round 2
Does the signal survive outside the estimation window?
Candidate 1 was tested against the random-walk Benchmark across 1,623 rolling pseudo-out-of-sample historical forecasts. The exercise asks two separate questions: whether Candidate 1 improves forecast accuracy, and whether its NQ pre-market signal remains economically persistent through time.
Historical forecasts
1,623
rolling pseudo-OOS sessions
MAE reduction
19.15%
Candidate 1 vs Benchmark
Positive NQ loading
100%
of rolling LAD fits
Peak → current loading
−22.22%
63-session smoothed beta
Signal stability
Persistent direction. Changing magnitude.
Current βNQ
0.90
Hover or tap to inspect the rolling coefficient.
Each coefficient comes from a rolling LAD estimation using information preceding the forecast observation. The NQ loading remained positive throughout the historical tournament, while its magnitude strengthened into 2022–23 and subsequently declined.
Forecast performance
Candidate 1 beat the Benchmark.
Historical MAE fell from 161.07 index points for the Benchmark to 130.22 for Candidate 1 — a reduction of 30.85 points, or 19.15%. Candidate 1 also recorded lower annual MAE in every calendar year represented in the tournament.
Interpretation
Attenuation is not disappearance.
The rolling NQ coefficient remained positive in every estimation window, but its magnitude was not constant. The evidence is therefore more consistent with a persistent, time-varying signal loading than with a fixed structural coefficient.
Evidence boundary.
Candidate 1 was selected from the frozen model tournament using historical
pseudo-out-of-sample performance. These 1,623 observations therefore
constitute model-selection evidence rather than an independent prospective
confirmation sample. Live prospective forecasts are tracked separately.
Research background
Why futures, and where the logic stops.
Four short reference pages separate the market definition, the instrument, the price-discovery evidence and the limits of what that evidence supports.
Engineering
Designed to run without me.
The model was first developed locally, then migrated into a cloud workflow. The production engine and post-close scorer have both been independently reproduced on a Linux cloud runner, so the process does not depend on my Mac being switched on.
DATA
Market-data pipeline
Programmatic futures data retrieval with explicit endpoint, timing and contract-roll handling.
Explore engineering →
CLOUD
Automated execution
Forecast generation and post-close scoring are designed to run remotely through version-controlled cloud workflows.
Explore engineering →
AUDIT
Reproducible records
Model versions, forecasts and scored outcomes are recorded separately to distinguish live evidence from retrospective tests.
Explore engineering →