Update README.md

README.md

@@ -26,11 +26,11 @@ A gap-free 1-minute OHLCV dataset for **BTCUSDT Binance USDⓈ-M Perpetual Futur
 covering five full calendar years: **2021-01-01 through 2025-12-31 (UTC)**.
 
 This repository contains **raw market bars only**. Feature engineering, aggregation,
-sample construction, normalisation, and temporal splitting belong to the downstream
+sample construction, normalization, and temporal splitting belong to the downstream
 [DiffQuant](https://github.com/YuriyKolesnikov/diffquant) pipeline, described below
 for reproducibility.
 
-Full codebase: [GitHub Repository](https://github.com/YuriyKolesnikov/diffquant)  
+Full codebase: [GitHub Repository](https://github.com/YuriyKolesnikov/diffquant)
 
 ---
 
@@ -244,16 +244,22 @@ cfg.data.preset          = "custom"
 cfg.data.feature_columns = ["close", "volume"]
 ```
 
-### Step 4 — Temporal splits (DiffQuant defaults)
+### Step 4 — Temporal splits
+
+The full dataset supports arbitrary split boundaries via `SplitConfig`.
+The primary DiffQuant experiment used the following non-overlapping splits:
 
 ```
-Train    : 2021-01-01 → 2025-03-31  (~4.25 years)
+Train    : 2024-01-01 → 2025-03-31  (15 months — intentionally recent)
 Val      : 2025-04-01 → 2025-06-30  (3 months)
-Test     : 2025-07-01 → 2025-09-30  (3 months)
-Backtest : 2025-10-01 → 2025-12-31  (3 months)
+Test     : 2025-07-01 → 2025-09-30  (3 months, out-of-sample)
+Backtest : 2025-10-01 → 2025-12-31  (3 months, final hold-out)
 ```
 
-Boundaries are fully configurable via `SplitConfig`.
+The training window is deliberately limited to 15 months rather than the full
+historical record. This keeps the training regime close to the evaluation periods
+and minimizes distribution shift. Extending to earlier data is the recommended
+first ablation and is straightforward via `SplitConfig.train_start`.
 
 ### Step 5 — Full pipeline one-liner
 
@@ -276,15 +282,15 @@ when the config changes (timeframe, preset, split boundaries, feature flags).
 ## Project context
 
 This dataset is the data foundation for **DiffQuant**, a research framework
-studying direct optimisation of trading objectives.
+studying direct optimization of trading objectives.
 
-Most ML trading systems suffer from a structural misalignment: models are trained
+Most ML trading systems face a structural misalignment: models are trained
 on proxy losses — MSE, cross-entropy, TD-error — while performance is measured in
-realised PnL, Sharpe ratio, and drawdown. DiffQuant studies what happens when this
-proxy is removed entirely: the full pipeline from raw features through a
-differentiable mark-to-market simulator to the Sharpe ratio is a single computation
-graph. `loss.backward()` optimises what the strategy actually earns, with
-transaction costs and slippage accounted for in every gradient update.
+realized PnL, Sharpe ratio, and drawdown. DiffQuant studies what happens when
+this proxy is removed entirely: the full pipeline from raw features through a
+differentiable mark-to-market simulator to the Sharpe ratio becomes a single
+computation graph. `loss.backward()` optimizes what the strategy actually earns,
+with transaction costs and slippage accounted for in every gradient update.
 
 **Key references:**
 
@@ -294,7 +300,7 @@ transaction costs and slippage accounted for in every gradient update.
 
 - Moody, J., Saffell, M. (2001). *Learning to Trade via Direct Reinforcement.*
   IEEE Transactions on Neural Networks, 12(4).
-  — original formulation of direct PnL optimisation as a training objective.
+  — original formulation of direct PnL optimization as a training objective.
 
 - Khubiev, K., Semenov, M., Podlipnova, I., Khubieva, D. (2026).
   *Finance-Grounded Optimization For Algorithmic Trading.*
@@ -326,4 +332,4 @@ transaction costs and slippage accounted for in every gradient update.
   publisher = {Hugging Face},
   url       = {https://huggingface.co/datasets/ResearchRL/diffquant-data},
 }
-```
+```