Switching-Cost-Regularised Fitted-Q Factor Rotation on a Point-in-Time S&P 500 Panel: Monthly Choice Among Momentum, Low-Volatility, Long-Term-Reversal (value proxy) and Short-Term-Reversal Long-Short Books (USEQ 1-DAY, dollar-neutral)

Derived by the Research Lead from Cordoni, Sancetta and Zhao, 'Dynamic Strategy Selection' (arXiv 2307.04754, Sections 2-4 and Supplement S.5): each month, pick ONE of several competing factor strategies with a fitted-Q (iterated ridge regression) policy whose reward penalises switching by a regularising cost larger than the real trading cost. The selection algorithm, state design (12-month factor trends + market-cycle dummies), reward (quadratic utility of net return), gamma=0.9, cap B=0.5, GCV ridge grid, NA=10 simulated action sequences and 120-month expanding initial window are taken from the paper. What I changed and why: the paper's actions are Kenneth French MOM/HML/RMW plus a CRSP low-vol factor; we hold no fundamentals (book value, operating profitability) and no CRSP, so I build the four actions as PRICE-ONLY long-short books from our own point-in-time S&P 500 daily panel: (1) MOM = 12-1 month return winners minus losers; (2) DEF = low 252-day realized volatility minus high volatility (the paper's DEF, Novy-Marx style); (3) LTR = 36-to-12-month long-term reversal (past losers minus winners), the standard price-only proxy for HML value (De Bondt-Thaler / Fama-French LT_Rev); (4) STR = 1-month short-term reversal (last month's losers minus winners), replacing RMW which cannot be built without fundamentals. Universe AS OF EACH month-end rebalance: S&P 500 members at that date via index_member()/universe_at() (former members included where bars exist), at least 400 sessions of bars, then the 150 most liquid by trailing 60-session median dollar volume (close x volume). Each factor book: equal-weight long its top quintile (30 names) and equal-weight short its bottom quintile (30 names) of the 150-name universe. The strategy computes all four factor books' daily returns internally every month from the panel (whether held or not), so the factor-return history used for the policy is point-in-time and available from the first month the 36-month LTR lookback is filled (~2000 with membership from 1996). Only the CHOSEN book is traded. Fills: decision on the last session close of the month, market orders filled at the next session's open (platform next_open_fill, supported). Risk control route (b): signal/rebalance exit, dollar-neutral, gross 100% of equity (50% long / 50% short, 1.67% equity per name), no single position above 2% of equity, so leverage 1x gross on the Reg-T 2x account; a long-short price factor at this gross runs roughly 8-15% annual volatility, inside the conservative 60% band. Dividends are not in PnL (both sides, roughly offsetting for a dollar-neutral book; flagged). Short borrow costs are not modelled; universe is restricted to the 150 most liquid large caps to keep hard-to-borrow names out. Benchmarks the analyst must see beside the policy (pre-registered controls, computed by the strategy as report-only series): the Greedy policy (gamma=0, one iteration) and each fixed factor held throughout the same OOS span, as in the paper's Table 1.