Fine-tuning adapts a general-purpose language model to a specific signal schema by training it on thousands of examples. In portfolio management, this turns a capable but poorly matched zero-shot model into a more useful signal generator. section 12.8
Measure whether the change helps: signal agreement with the oracle rose from 81% to 91% across two iterations, while the lead over other signal modes widened. Changing one component at a time helps trace whether improvements reach downstream performance. section 12.8
Table 12.2 Five modes on the test window, averaged over all episodes
|
Mode
|
Sharpe
|
Annualized
Return
|
Annualized
Volatility
|
Maximum Drawdown
|
Turnover
|
Final Value
|
|
Buy-and-hold
SPY
|
0.77
|
26.6%
|
34.4%
|
−12.33%
|
1.00
|
1.0432
|
|
Equal-weight
|
0.91
|
16.0%
|
17.5%
|
−6.34%
|
1.16
|
1.0216
|
|
RL, no news
|
1.41
|
40.9%
|
29.1%
|
−9.76%
|
4.07
|
1.0571
|
|
RL + naive
signal
|
2.52
|
67.1%
|
26.6%
|
−7.06%
|
6.09
|
1.0876
|
|
RL + base
LLM signal
|
2.62
|
66.7%
|
25.5%
|
−7.07%
|
6.49
|
1.0871
|
|
RL +
fine-tuned LLM signal
|
8.47
|
169.5%
|
20.0%
|
−2.18%
|
8.96
|
1.2251
|
Compare test-window results across portfolio strategies; the fine-tuned signal mode reports a Sharpe ratio of 8.47, versus 2.62 for the base LLM signal. section 12.8



















































