DDPG's gain from putting sentiment into the reward is worth pursuing, though far too thin to trade on. Among the five agents Singh, Sharma and Pant train, only DDPG improves clearly when sentiment moves from the state into the reward: return rises from 0.46 to 0.70. SAC gains 0.05; the other three lose ground. And the added reward never checks what the agent holds. Those facts limit what we can take from the headline 0.70 annual return and 2.75 Sharpe.

Inside the experiment

The authors start from the premise that price indicators miss event-driven moves, and that reading headlines should help an agent trade. They collect 1 to 10 Google RSS headlines per stock-day for five Indian large caps: Reliance, Adani Ports, Bharti Airtel, HDFC Bank and ICICI Bank. FinBERT scores each headline. Its top label receives a sign (+1 positive, -1 negative, 0 neutral), which is multiplied by the score. Averaging these values per stock-day gives an "inclination score" θ between -1 and 1. A day without headlines receives θ = 0.

Daily prices come from Yahoo Finance and span January 2019 to June 2024. Seven technical indicators enter the state: MACD, Bollinger Bands, RSI-30, CCI-30, DX-30, 30 and 60-day SMAs. The dataset also includes CBOE VIX and a turbulence index. The five actor-critic agents trade a ₹1 million portfolio from June 2023 to June 2024. A2C and PPO use pure policy-gradient methods; DDPG, TD3 and SAC also use Q-learning critics.

The authors run each agent in three configurations. Phase A excludes sentiment and rewards changes in portfolio value. Partial infusion adds θ to the state. "Deep infusion" includes θ in both the state and the reward.

PPO leads without sentiment, returning 0.45 with Sharpe 2.50 and drawdown -0.06. Partial infusion puts TD3 (0.63, 2.46) and A2C (0.61, 2.36) ahead, while PPO falls to 0.26 and 1.32. DDPG leads under deep infusion: return 0.70, Sharpe 2.75, drawdown -0.07 and "stability" 0.95. Its no-sentiment return and Sharpe were 0.31 and 2.09. The authors define stability as one minus the standard deviation of portfolio growth divided by its mean.

Does the reward add anything beyond the state?

For two of the five agents, it does. Moving from partial to deep infusion changes return and Sharpe as follows:

SAC's Sharpe gains just 0.01.

The abstract claims that Q-learning agents improve "significantly and consistently" under deep infusion. The conclusion makes a narrower, more candid claim. It singles out DDPG and SAC, acknowledging TD3's decline and attributing it to more conservative value estimation. In the results section, the explanation is that sentiment increased reward variability and drove TD3's double Q-learning to under-explore.

That explanation would carry more weight if the rankings held steady elsewhere. They do not. Adding the sentiment score to the state takes PPO from best to worst. A2C returns 0.16, 0.61, 0.26 across the three phases. We found no seeds, repeated runs or confidence intervals, leaving no way to distinguish these swings from seed noise. Only DDPG improves on both return and Sharpe throughout: 0.31, 0.46, 0.70, with Sharpe 2.09, 2.15, 2.75. SAC's return rises too, from 0.34 to 0.41 to 0.46, while its Sharpe falls from 2.14 to 2.02. As far as we can tell, each phase of DDPG's progression comes from one run.

A bonus detached from the position

Equation 5 multiplies the price change Δρ of each stock by its θ. Equation 6 averages those products across the five stocks. Equation 7 then adds that average to the change in portfolio value. The bonus contains no holdings term.

Prices agreeing with headlines earn the bonus, regardless of whether the agent owns the stock or stays flat. We found no market impact model in the paper. With a price-taking simulator, neither Δρ nor θ is under the agent's control, so every policy receives the same bonus. The term changes the critic's targets and makes them more variable. No particular trade earns it.

The 0.70 therefore shows that DDPG trained differently with a noisier target. It offers weak evidence that trading in line with sentiment makes money. There is also a scale issue: Δρ is in rupees, rather than returns. A higher-priced stock contributes more to the bonus than a cheaper stock making the same percentage move. Multiplying θ by the position's return would serve the stated intent. It would also produce a different reward from the one tested in the paper.

Implementation requires guesses

Some definitions cannot be implemented literally. Equation 3 adds cash to the number of shares held to calculate asset value; an implementer has to read that as shares times price. The annual-return definition is ((P_A − P_0)/P_0 − 1)×100. This would make any gain below 100% negative, whereas the tables report 0.16 to 0.70. Max drawdown is given as (P_H − P_L)/P_L. That expression is positive and takes no account of whether the trough follows the peak, yet the tables report negative values. We found no stated risk-free rate for Sharpe.

Buy, sell or hold is the stated action space, although DDPG, TD3 and SAC take continuous actions. We found no account of position sizing. VIX brings a US index into trading NSE or BSE names. The paper does not state which exchange supplies its prices, as far as we could find, although the architecture figure uses an NSE schema.

Assigning θ = 0 makes a headline-less day equivalent to a neutral day. Yet the merged sentiment file contains 4,843 rows. At 252 sessions, five stocks across the paper's 5-year news timeline would yield about 6,300 stock-days. The 4,843 rows fall short even of that 5-year count, despite the stated practice of assigning 0 to days without headlines. The paper leaves the gap unexplained.

We found no training window, hyperparameters or network sizes. We encountered the same kind of omission in an actor-critic allocator. Table 1 gives market caps as of September 2024, after the test window ends. Choosing the five names at that date would introduce look-ahead into the universe. There is also a small conflict between text and table that matters when checking the work: the paper describes A2C's -0.1293 as the worst no-sentiment drawdown, while Table 2 gives TD3 -0.15.

The missing passive comparison

A single window, June 2023 to June 2024, produces DDPG's 0.70 return, 2.75 Sharpe and -0.07 drawdown. We could find neither a cost assumption nor a buy-and-hold or equal-weight comparison for those same five stocks. A Sharpe above 2 on five large caps could largely reflect market exposure without that comparison. Partial infusion raises volatility for all five agents. Deep infusion raises it for four of five, with TD3 falling from 0.19 to 0.17. DDPG's volatility rises from 0.14 to 0.19 to 0.20; A2C moves from 0.15 to 0.21 in both sentiment phases.

We could not backtest this. We lack Indian-listed price histories for the run, and our news history does not extend to the January 2019 start. US stocks would substitute a different market and different news flow. We therefore did not make that substitution.

We would take the result seriously if DDPG kept its advantage over partial infusion across ten seeds, net of NSE costs and against equal-weight, with a rewritten bonus that depends on holdings.