> Research status: This is an observational report on 351 active STT/current runs queried on 2026-08-11. Every August 11 row carried a zero next-day return because the bar was still incomplete, so the report uses completed daily data from July 12 through August 10. It does not forecast returns, establish a causal window effect, or report live-account PnL.

The question

Does a 32, 64, or 128-bar context window change STT performance over the most recent 30 completed days?

The question has two separate measures:

1. How does the 30-day return distribution change across individual models? 2. What happens after we average target positions by symbol and window, apply the one-day execution lag, and charge costs once to the netted turnover?

The first measure describes a model population. The second describes the position-level portfolio contract used for an operating ensemble. We keep them separate.

Short answer

• The 30-day position-level returns for 32, 64, and 128 bars were +0.25%, +0.22%, and +0.27%. Their endpoints were close.
• Individual 128-bar models had the highest median return (+0.15%) and the highest positive-model share (55.9%). The same window portfolio also had the highest annualized volatility (1.80%) and the deepest drawdown (-0.40%).
• The 64-bar portfolio finished with the smallest return, but it also had the lowest annualized volatility (1.04%) and drawdown (-0.32%). A 30-day annualized Sharpe ratio is too unstable to choose a winner.
• Spearman correlation between log2(window) and 30-day return was -0.01. Correlation between window length and total turnover was -0.41. Longer windows changed less often in this cohort; they did not show a simple return advantage.

Snapshot scope

A common market path for model and portfolio comparisons

Some model rows recorded different asset_return values for the same symbol and date because their backfills came from different data versions. Sixty-four of 390 symbol-date observations (16.4%) had more than one value. This report uses the modal value among active STT rows as the canonical market return, so every window faces the same market path.

We recalculate individual-model distributions and window portfolios under the same timing and cost contract:

ensemble_target[t] = mean(target_position[t], same symbol + same window)
realized_position[t] = ensemble_target[t - 1]
turnover[t] = abs(ensemble_target[t] - ensemble_target[t - 1])
net_log_return[t] = realized_position[t] * log(1 + canonical_asset_return[t])
                    - turnover[t] * execution_unit_cost[t]
position_return[t] = exp(net_log_return[t]) - 1
portfolio_return[t] = mean(symbol_position_return[t], 13 symbols)

execution_unit_cost uses each current run's fee and slippage settings with the current build_strategy_curve volatility-dependent slippage formula. The daily table does not retain liquidity stress, so the calculation fixes it at the neutral value of 1.0.

The dashboard below renders the verified aggregate snapshot in the browser. We do not publish raw model rows, CSV files, or chart image artifacts.

Reading the charts

128 bars followed a different path, not a decisive better endpoint

The 128-bar group led the individual-model median at +0.15% and the positive-model share at 55.9%. Portfolio endpoints still clustered tightly: +0.25% for 32 bars, +0.22% for 64 bars, and +0.27% for 128 bars.

The 128-bar portfolio fell to about -0.35% during mid-July before recovering in early August. It finished with a -0.40% maximum drawdown and 1.80% annualized volatility. Its endpoint ranked first, but the route was not smoother.

The 64-bar portfolio reached higher interim values, then closed at +0.22%. Its 1.04% volatility and -0.32% drawdown were the lowest of the three. The sample is too short to turn that observation into a risk-adjusted ranking.

Window length tracked turnover, not recent return

The correlation matrix gives ρ = -0.01 for log2(window) and 30-day return. Window length alone did not explain the cross-sectional return order in this active snapshot.

The relationship with turnover was clearer. The window-level total turnover fell from 1.59x at 32 bars to 1.22x at 64 bars and 1.15x at 128 bars, while the Spearman correlation was -0.41. Longer input histories coincided with less frequent signal changes.

Average absolute position and annualized volatility had a +0.76 correlation. Return and average absolute position had a +0.02 correlation. Larger exposures raised observed volatility here; they did not explain return differences.

Backbone and training generation confound a window-only claim

The backbone-by-window cells do not move in one direction. FASCL had median returns of -0.08%, +0.77%, and +0.14% across 32, 64, and 128 bars. Franceschi showed +0.77%, -0.36%, and +0.44%. TS2Vec reached +0.17% at 128 bars while its 32- and 64-bar medians were close to zero or negative.

The active cohort also mixes 105 models trained with a 10% train fraction and 246 trained with a 30% train fraction. Within 32 bars, the 10% generation averaged +0.63% and the 30% generation was nearly flat. Within 64 bars, the direction reversed.

These models do not isolate window size. Backbone, training fraction, and model generation all move with the window groups.

The 5Y gate score and this 30-day slice answer different questions

Spearman correlation between the 5Y gate Sharpe and recent 30-day return was +0.03. That does not invalidate the gate. It says that, across this active STT cohort and this short cross-section, the historical gate score did not rank the next 30 days of returns.

The gate summarizes a much longer selection period. This report observes one completed month. They should not be treated as interchangeable metrics.

Observed, unproven, next

• Observed: all three window portfolios posted positive 30-day returns between +0.22% and +0.27%.
• Observed: 128 bars led the model-level median and positive share, while its portfolio had the highest volatility and deepest drawdown.
• Observed: longer windows coincided with lower turnover in this snapshot.
• Unproven: increasing context length causes better return.
• Unproven: these 30-day results will persist over later 30-day or 90-day windows, or in live-account PnL.
• Next: fix matched cohorts by backbone, symbol, train fraction, and seed before training, then repeat this aggregation using only completed forward bars.

Limits

• The report applies the current active model set to the preceding 30 days, so selection bias remains.
• Models from the same family, symbol, and training generation are not independent observations.
• Funding, unfilled orders, latency, account leverage, and account-level risk limits are outside this calculation.
• The canonical modal market return makes the window comparison consistent; it does not reconstruct the original data version of every historical row.
• STT's random sliding-window train/test split and full-period gate are not independent out-of-sample evidence. This report does not turn one month of observations into that evidence.

This is a research record, not investment advice.