Update README.md

README.md

@@ -222,8 +222,8 @@ Applied channel-by-channel after aggregation. The first bar is always dropped
 | Channel | Transformation |
 |---|---|
 | open, high, low, close | `log(price_t / close_{t-1})` — log-return vs previous bar close |
-| volume | `volume_t / global_mean(volume)` — ratio to mean of the full aggregated series |
-| num_trades | `num_trades_t / global_mean(num_trades)` — same |
+| volume | `log(volume_t / rolling_mean(volume, window) + eps)` — relative intensity |
+| num_trades | `log(num_trades_t / rolling_mean(num_trades, window) + eps)` — same |
 | typical_price (optional) | `log(((H+L+C)/3)_t / close_{t-1})` |
 | time features (optional) | `[sin_hour, cos_hour, sin_dow, cos_dow]` — cyclic UTC encoding |
 
@@ -274,13 +274,15 @@ 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:
-
-> In standard ML trading pipelines, models are trained on proxy objectives —
-> MSE for price prediction, TD-error for RL — evaluated indirectly through
-> downstream trading logic. DiffQuant studies a tighter formulation: position
-> generation, transaction costs, and portfolio path interact directly with the
-> Sharpe ratio as the training objective through a differentiable simulator.
+studying direct optimisation of trading objectives.
+
+Most ML trading systems suffer from 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.
 
 **Key references:**