Core · 7
Returns, volatility, volume, range, and close location preserve the original transparent baseline.
MODEL 1 · STEP 3Build
Build
the features.
Turn each completed ASTS session into a 23-feature snapshot using only stock and market history available at the end of session t.
MARKET CONTEXT · OUTSIDE MODEL FREEZEASTS
These readings update independently. They are not model inputs, deployment evidence, or a trading recommendation.
- Latest price
- $62.31 · fallback
- 20D momentum
- -13.4% Close-to-close · 20 completed sessions
- RSI 14
- 47.2 neutral · Wilder method
- Relative volume
- 1.0× Versus the prior 20 completed sessions
TECHNICALS THROUGH · TIMESTAMPED FALLBACK
What could we measure at close t?
Features are facts available when the prediction is formed. For five development runs, we measure ASTS behavior, trailing empirical probabilities, and synchronized market context—without letting the future label leak backward.
INPUTmarket_v2
TRANSFORM23 trailing transforms
OUTPUTasts_features_v2
THREE FEATURE FAMILIES
Expand the hypothesis without changing the clock.
The seven-feature core remains intact. Probability-state and market-context families expand the comparison while every run keeps a common eligible sample.
Probability · 7
Trailing up rates, smoothed five-session outcomes, percentiles, and tail frequencies describe recent state.
Market · 9
SPY trend and volatility, QQQ/IWM excess returns, breadth, and risk-on share add synchronized context.
FEATURE CONTRACTasts_features_v2
FROZEN / V2Features23 · 3 families
Warmup80 sessions maximum
Model rows1,249 common rows
Labeled through22 Jul 2026
FEATUREWINDOWFORMULA
return_1d1 prior sessionclose[t] / close[t−1] − 1
return_5d5 prior sessionsclose[t] / close[t−5] − 1
return_20d20 prior sessionsclose[t] / close[t−20] − 1
volatility_20d20 sessionspopulation stdev of returns through t
volume_ratio_20d20 sessionsvolume[t] / mean(volume[t−19:t])
intraday_range_pctSession t(high[t] − low[t]) / close[t]
close_locationSession t(close − low) / (high − low); 0.5 if zero range
up_day_share_20d20 sessionsshare of positive one-session returns
up_day_share_60d60 sessionsshare of positive one-session returns
p_up5_stride_60d60 sessionsBeta(2,2)-smoothed non-overlap up share
return_1d_ecdf_60d61 sessionssmoothed percentile of today’s return
abs_return_tail_60d61 sessionssmoothed absolute-return tail frequency
volatility_20d_ecdf_60d80 sessionsvolatility percentile versus prior values
volume_ratio_ecdf_60d80 sessionsvolume-ratio percentile versus prior values
spy_return_5d5 sessionsraw SPY close return
spy_return_20d20 sessionsraw SPY close return
qqq_excess_5d5 sessionsQQQ return minus SPY return
qqq_excess_20d20 sessionsQQQ return minus SPY return
iwm_excess_5d5 sessionsIWM return minus SPY return
iwm_excess_20d20 sessionsIWM return minus SPY return
spy_volatility_ecdf_60d80 sessionsSPY volatility percentile
market_breadth_up_20d20 sessionsmean up-day share of SPY, QQQ, IWM
risk_on_share_20d20 sessionsshare QQQ and IWM both beat SPY
POST-FREEZE TECHNICAL CHALLENGER
RESEARCH ONLYProbability error fell, but uncertainty still crosses zero.
A fixed benchmark adds 20-session momentum, Wilder RSI 14 with overbought and oversold flags, prior-20 relative volume, and price-volume confirmation. It is scored on the last 252 eligible labels without changing this model's frozen nine-gate decision.
- Brier error
- 0.258945 → 0.256266 Lower is better
- Brier gain
- +0.002678 95% paired interval −0.001732 to +0.006812
- 0.5 hit rate
- 45.6% → 48.8% Threshold accuracy alone cannot approve a model
- Challenger AUC
- 0.451 Below the 0.50 no-skill line
Promotion rule: not passed. The Brier confidence interval includes zero and discrimination remains below 0.50, so the readings stay a live context layer while genuinely new sessions accumulate.
Open formulas and audit filesSEAL THE FUTURE
Past builds X. Future builds y.
Session t may use its completed bar. Sessions t + 1 through t + 5 stay outside the feature matrix.
t − 80Longest lookback
t − 60State history
t − 20Core history
tFeatures
t + 1…4Sealed
t + 5Label
FEATURE BLOCK / X
Only trailing information.
Returns, volatility, volume ratio, range, and close location are joined by trailing probability-state and ETF context features; all are finalized by the decision boundary at session t.
LABEL BLOCK / Y
Kept physically separate.
Forward return and direction_5d are diagnostics for supervised learning. They are never passed to the classifier as features.
STEP 3 LEAKAGE GATE
Ten checks before training.
A high score cannot rescue a feature matrix that already knows the answer.
- 01
Every one of the 23 feature windows ends by the completed close at session t.
- 02
All five runs use the same 1,254-row common eligibility intersection.
- 03
The longest probability and market transforms receive an 80-session warmup.
- 04
The final five sessions stay feature-ready but unlabeled.
- 05
Trailing five-session outcome frequencies only use outcomes already realized by t.
- 06
SPY, QQQ, and IWM inputs join on the same ASTS session date.
- 07
Forward return and direction_5d remain in the label block, never the feature block.
- 08
Missing inputs make a row ineligible; they are not silently forward-filled.
- 09
Scaling, calibration, and model selection are deferred to Step 4 folds.
- 10
Trading-session offsets follow the exchange calendar, not calendar days.
Do not standardize or choose among feature families on the full dataset. Step 4 fits preprocessing, tuning, calibration, and selection inside the training chronology.
STEP 3 OUTPUTasts_features_v2
Twenty-three features, a common 1,254-row eligibility window, five unlabeled tail rows, and 1,249 model-ready examples.
SOURCE NOTES
The feature workbook inherits the four-asset frozen Step 2 market dataset. Raw daily bars and their source fingerprints remain documented there.
NEXTStep 4 · Train through time
Walk-forward validation comes next.