- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes DualIPO, a system that decides how much to allocate to each stock conforming a portfolio. It is composed by two networks. The first is a diffusion model, which generates many plausible future returns trajectories for each of the stock in the portfolio. The generated paths then are summarized into a few key statistics (average return, spread in the outcomes, etc.). The second component of the system is a neural network that takes as an input these statistics and outputs the corresponding portfolio weights. These weights are always non-negative and sum to one; meaning no short-selling or borrowing is allowed.
Training has two parts. First, the diffusion model trains on its own to just produce forecasting results. Then, both parts are trained together on the actual investment outcome. Training alternates between the two parts: in each step the forecaster is updated first, and the allocator is then updated using the forecaster's new scenarios. This pushes the forecaster to generate scenarios that lead to better investment choices.
The main claim of the paper is that training the proposed system on investment performance, rather than just on forecasting accuracy, leads to better portfolio allocations. The method is then tested on the 50 and 100 largest stocks of the US S&P 500 and the Chinese CSI 300, where the authors report the best Sharpe ratio in all four settings and compare it against some baselines: an equal-weight portfolio, a momentum strategy, classical mean–variance optimization, and two diffusion-based methods (STABLE and Diffusion-DFL).
Critical strengths
- The experiments are careful about leaking information from the future , which is a common problem in these approaches. I did not find that it happens here.
- The main idea makes sense here. Instead of solving an optimization problem to build. portfolio every time, you just take a neural network which is trained to learn to turn forecasting scenarios into portfolio weights. This is simpler and also faster once it has been trained. Also, training both the forecaster and allocator networks together on the final investment outcomes is the way to go here.
Critical weaknesses
This paper should be rejected because: (1) the testing period is too short to tell methods apart; (2) DualPO barely beats a simple equal-weight portfolio, while the other baselines look poorly tuned; (3) there is no proof that the diffusion model or the uncertainty it gives is what drives the results.
Reasons for rejection
-
The test period is only around six months (which is 120 trading days), and the results come from 3 seeds without any reported variance. Note the following: a Sharpe ratio is the mean of daily returns divided by their standard deviation, and estimating it over so few days can be very noisy. The daily Sharpe ratio is . The standard deviation of the estimated mean over days is , so the standard deviation of the estimated Sharpe ratio is . After annualizing, this becomes , which is about for . This means that a strategy with a true Sharpe ratio of, say, could give anything between and over this period of time just by pure chance. The gaps of vs against equal weight on S&P-100, or vs against the frozen ablation, are therefore clearly within this noise and not meaningful.
-
One of the paper’s central claims is not really supported: nothing is showing that the forecaster will produce meaningful uncertainty after training. More concretely, the method is presented as using the predictive uncertainty given by the forecaster multiple outputs, but then, in the second training phase, the forecasting (denoising) loss is heavily down-weighted, making the forecaster being rewarded by just producing scenarios that make the allocator network output weights that are profitable in the training period rather than producing realistic forecasts. This makes overfitting to the training period likely. Furthermore, forecast quality after this training stage is never measured.
Questions for authors
- Could you put in the paper the per-seed means and standard deviations as well as to evaluate all methods on a longer test period? This would address the first weakness above.
- How were the baselines tuned? Did they get a tuning budget similar to DualPO's one?