Benchmark hardcoded to SPY breaks reflection alpha for non-US tickers

[adityarai7297]

Problem

_fetch_returns() in tradingagents/graph/trading_graph.py:205 always pulls SPY as the benchmark when computing alpha for the reflection layer:

spy = yf.Ticker("SPY").history(start=trade_date, end=end_str)
...
alpha = raw - spy_ret

The reflection log line in tradingagents/graph/reflection.py:48 hardcodes the same assumption:

f"Alpha vs SPY: {alpha_return:+.1%}\n\n"

This is fine for US-listed tickers but produces misleading alpha for international ones. For example, analyzing RELIANCE.NS (NSE) against SPY mixes a single-stock INR return with a USD index return — the resulting "alpha" reflects FX and regional macro drift more than stock-specific performance. Same issue for .T (Japan), .HK (Hong Kong), .L (London), .TO (Toronto), etc.

The framework already advertises international support — cli/utils.py:11 literally lists CNC.TO, 7203.T, 0700.HK as ticker examples, and agent_utils.py:37-43 prompts agents to preserve exchange suffixes — so the reflection layer is the odd component out.

Proposed solution

Make the benchmark configurable in default_config.py, with optional auto-detection from the ticker's exchange suffix. Two shapes worth discussing:

Option A — single override:

"benchmark_ticker": "SPY"

Option B — suffix map with auto-detect (preferred):

"benchmark_ticker": None,  # None = auto-detect via suffix map
"benchmark_map": {
    ".NS": "^NSEI",    # Nifty 50
    ".BO": "^BSESN",   # Sensex
    ".T":  "^N225",    # Nikkei 225
    ".HK": "^HSI",     # Hang Seng
    ".L":  "^FTSE",    # FTSE 100
    ".TO": "^GSPTSE",  # TSX Composite
    "":    "SPY",      # default for US-listed
}

Option B keeps the zero-config experience for US users while making non-US analyses meaningful out of the box. An explicit benchmark_ticker value would still win over the map for advanced users.

reflection.py would then format f"Alpha vs {benchmark}: ..." instead of hardcoding SPY.

Scope

Strictly the benchmark. Out of scope for this issue (each its own follow-up):

• Currency normalization for TraderAction price targets
• News source localization (Moneycontrol, Economic Times, etc.)
• Region-specific insider/disclosure feeds (SEBI for India, etc.)

Willing to PR

Happy to send a PR against main once a maintainer confirms which shape (A vs B) is preferred. Existing tests in tests/test_memory_log.py already cover the _fetch_returns path and would be extended.