The forecasts supply most of the trading gain in Annalakshmi and Thangapriya's ablation. Taken at face value, the paper gives drift adaptation, the feature behind the method's name, credit for 1.3 ROI points (24.8% to 26.1%).
What the SVM-XGBoost blend gives the agent
DACH-DDPG (Drift-Aware Continuous Hybrid DDPG) feeds a blended forecast into a DDPG agent. It trains an RBF-kernel SVM and an XGBoost model on daily OHLCV together with RSI, MACD, EMA, SMA and volatility indicators. Both models predict direction or return. Validation performance determines the weights in a convex combination of their outputs. This forecast joins the features, current holdings and drift information in the agent's state. An MLP actor, with 2 to 3 hidden layers of 128 to 256 ReLU units, produces a position in [-1, 1].
The reward subtracts a transaction-cost term and a lambda-weighted risk penalty from the change in portfolio value. Rolling volatility or drawdown supplies that penalty. Feeding the forecast into the state gives the policy a directional prior that it would otherwise have to learn from raw prices.
Adaptation happens during deployment: the system compares rolling performance with a reference level. A sufficiently large shortfall prompts retraining of both predictors on recent data and recalibration of the blend weights. The paper also offers a test of changes in the feature distribution as an optional alternative trigger.
For data spanning 2018 to 2025, the paper's headline results are 26.1% ROI, Sharpe 1.51 and maximum drawdown -11.8%. It reports PPO at 19.3%, 1.08 and -16.7% on those measures, DQN at 16.2%, 0.91 and -18.4%, and the blend traded by itself at 23.6%, 1.32 and -14.2%. Classification accuracy for the blend reaches 79.8%, compared with 71.2% for the SVM and 74.6% for XGBoost. What the DDPG rows classify to achieve their 76.4% and 83.5% accuracy is left unspecified.
Forecasts contribute 4.4 of the 5.7-point ROI gain over plain DDPG
Plain DDPG, without a forecast in its state, earns 20.4% in the ablation, with Sharpe 1.12 and drawdown -16.3%. XGBoost alone takes ROI to 24.1% and Sharpe to 1.36. With the full blend and no drift adaptation, the figures reach 24.8%, 1.41 and -13.1%. Switching adaptation on produces 26.1%, 1.51 and -11.8%.
Forecast inputs therefore account for 4.4 points of the 5.7-point ROI improvement over plain DDPG. Adaptation contributes 1.3. The corresponding Sharpe contributions are 0.29 and 0.10. XGBoost supplies most of the forecast improvement; including the SVM adds 0.7 points beyond XGBoost alone. The conclusion describes the integration of drift detection and model adaptation as "a key feature." The table assigns most of the gain to the forecasts.
Does the trigger earn more than a weekly refit?
Training ran "with continuous retraining scheduled weekly to incorporate new market data," according to the setup. Yet the drift table records three events: SVM/XGBoost retraining in 2020, a blend-weight adjustment in 2022, and incorporation of recent data in 2024. Weekly refitting across 2018 to 2025 would mean hundreds of refits. The three recorded events are hard to reconcile with that schedule. We could not establish which regime generated the headline figures.
This uncertainty carries into the ablation. We did not find an account of how the "without drift adaptation" row handles retraining. If that configuration never refits, its 1.3-point deficit measures the value of retraining in any form. A simple calendar schedule might recover all of it. If it continues to refit weekly, the 1.3 points can be attributed to the trigger, as the paper's name implies.
A matched comparison would resolve this: refit the same models on the same calendar, with and without the trigger. We did not find that comparison. Nor did we find seeds, standard deviations or confidence intervals. The 1.3 points consequently appear to come from a single run of each configuration. RL results vary with the seed.
The Sharpe figures disagree
Table 1 gives the proposed model 83.5% accuracy, RMSE 0.128 and Sharpe 1.34. In Table 2, its Sharpe is 1.51. The standalone blend also changes between tables, from 1.21 in Table 1 to 1.32 in Table 2.
The Figure 3 discussion describes "a peak ROI of approximately 31% around 2024." Whether the tables' 26.1% represents the end of that curve is unstated. The training table also records 26.1% ROI at episode-500, matching the headline exactly. This raises the question of whether the headline reports training performance. Every row of Table 2 gives identical values for ROI and cumulative return. Those equalities fit a single holding period, rather than an annualized return across several years.
Even the universe shifts.
The methodology specifies "NIFTY 50 and S&P 500 constituents"; elsewhere, the setup includes the "NASDAQ composite" index. We did not find a stock count or split dates. A test window and a cost level in basis points were also missing from what we found, although the paper says it includes costs. The setup promises an LSTM baseline and buy-and-hold, neither of which appears in the results tables.
An accuracy of 83.5% on classification using what the paper calls daily data would alone warrant a leakage audit. The paper says it maintains chronological integrity. We could not locate the window used to fit the min-max bounds.
Our substitute trades the blend on US large caps
Our strategy substitutes a different universe and neither tests nor verifies the published figures. We trade US equities daily because we do not have NIFTY 50 data. To approximate S&P 500 membership, we screen annually for the 500 largest names by capitalization. This differs from official index membership. Results reported in the paper for either original universe therefore do not transfer to our screen.
We lag every input: a signal formed at the close trades in the following session. Only trailing data determine normalization bounds, imputation means and blend weights.
Compute is the larger departure. We do not have the resources to train DDPG at the paper's stated scale. Our strategy instead trades the signal from the SVM and XGBoost blend, the paper's own classical baseline. The authors report 23.6% ROI and Sharpe 1.32 for that configuration. Any return our version earns tests the predictor layer; DACH-DDPG's claim remains untested. The predictor layer does allow a cheap version of the comparison we did not find in the paper: refitting weekly versus refitting only when the drift trigger fires.