custom_index.py / fo_effect.py — Custom Indices (Leverage Study, F&O Effect)
Research question (2026-09-27): can leverage generate extra alpha?
backend/nseeod/custom_index.py builds an index of all NSE F&O stocks:
- equal-weighted;
- held through their near-month futures;
- rolled at every monthly expiry;
- base 1000, from the first stock-futures expiry (Nov 2001).
It is built at 1x, 2x and 3x (configurable). Results live in their own tables (ci_master, ci_level,
ci_holding), never in scr_master, scr_nseeq_eod or the NSE-index idx_* tables. They are
shown on the Leverage Study page.
Method
At each monthly expiry R(k) (the expiry-day close):
| Step | Rule |
|---|---|
| Universe | Stocks whose future expiring R(k+1) trades on R(k): the F&O list as it was then, so survivorship-free. Stocks added to F&O join from the next cycle; stocks removed leave at expiry, as NSE runs its contracts. |
| Position | q_i = (L · NAV / N) / F_i(R(k)), so the notional is L × NAV, split equally |
| Daily NAV | NAV(t) = NAV(t−1) + Σ q_i·(F_i(t) − F_i(t−1)) + NAV(t−1)·r·days/365 |
| Roll | Close at the expiring contract's close on R(k+1), open the next contract at its close; leverage resets to L |
| Costs | cost_bps (default 5) per side on every notional traded |
| Ruin | NAV ≤ 0 stops the series |
Rebalance dates are the actual monthly expiries, taken from the stock futures' own last trading days
(rebalance_dates(), 299 so far; see Validation): holiday-shifted, and on Tuesday from Sep 2025.
Collateral rate r (decided 2026-09-27: implied): each expiry,
r = (F_NIFTY / S_NIFTY − 1) · 365 / days + NIFTY dividend yield.
- F is the NIFTY future expiring R(k+1) and S is the NIFTY spot close.
- The future's premium is funding cost minus dividends, so the dividend yield (
scr_indexpepb, from 2012) is added back. Before 2012 the 2012–26 average of 1.3% is used. - The rate is clipped to 0–20%.
- It is noisy, and below true money-market rates in some years (≈1.4% in 2005–06, 2.4% in 2020), because NIFTY futures traded near or below spot. It is still the market's own price of funding.
- The rate applied each day is stored in
ci_level.cash_rate.--cash-rate 6.5gives a fixed rate instead.
Splits / bonuses inside a cycle: NSE adjusts running futures (price down, quantity up; RELIANCE's
Sep-2017 future went 1,646.95 → 821.25 on the 1:1 bonus). Each contract's earlier prices are scaled by
the stock's scr_corp_action factor, only when the contract's own jump confirms it: within ±18% in log terms,
or, for a big split (factor and jump both below 0.8), within ±50%, since a split can land on a crash day.
Ordinary dividends need nothing, since futures already price them.
Positions are fractional (it's an index). A lot-rounded variant (lotsize) is possible later.
Measuring "extra alpha"
- CAGR, volatility, Sharpe, max drawdown, Calmar, worst month, longest drawdown per series.
- CAGR − L × 1x CAGR: the volatility drag, i.e. what leverage lost to compounding and costs.
- Growth-optimal (Kelly) leverage of the 1x series,
(mean excess return) / variance: the leverage that would have maximised compound growth. - Not the answer: regression "alpha vs 1x". With futures leverage it is mechanical, about −(L−1) × r: every series earns interest on the same collateral. It's stored but not interpreted.
Results (local, 2026-09-27): Nov 2001 → Sep 2026, base 1000
| Series | Final NAV | CAGR | Volatility | Sharpe | Max drawdown | Worst month | CAGR − L × 1x |
|---|---|---|---|---|---|---|---|
| 1x | 16,259 | 11.9% | 24.7% | 0.37 | −71.9% | −31.1% | — |
| 2x | 7,074 | 8.2% | 53.8% | 0.33 | −97.1% | −73.2% | −15.6 pts |
| 3x | 0: wiped out 2008-10-23 (it had reached 35,611 by end-2007) | — | — | — | −100% | — | — |
Growth-optimal (Kelly) leverage of the 1x series: 1.50x.
Answer to the research question, for this universe and period: leverage did not add alpha.
- Return per unit of risk falls slightly with leverage (Sharpe 0.37 → 0.33).
- Compounding punishes it: 2x ended at less than half of 1x's value after a −97% drawdown in 2008.
- 3x was wiped out in the 2008 crash.
- The best constant leverage in hindsight is only ~1.5x, and that assumes the 1x volatility and drawdowns are bearable.
- The one exception is a strong trending stretch: 2001–2007, when 2x and 3x did far better than 1x before giving it all back.
Universe size per cycle: 31 stocks (Nov 2001) → 200+ (2007 onward, 133–228 by cycle).
Validation (engine checks)
Full range, Nov 2001 → Sep 2026:
| Check | Result |
|---|---|
| Leverage identity, first day of each cycle | exact (0 error) on all 298 cycles (2x) and all 83 cycles 3x lived |
| 1x futures + collateral vs an equal-weight spot total-return series (same members and expiries) | daily correlation 0.995. The futures route trailed spot TR by 1.7% a year: ~1.2% roll costs (5 bps × 2 sides × 12), the rest the stock-futures premium over the NIFTY-implied rate |
| Splits / bonuses inside a cycle | 284 scaled, 0 left unconfirmed. JINDALSTEL's 1:5 split on the 2008-01-21 crash day (future 0.14× vs factor 0.2) counts as confirmed under the "big split" rule |
| Two contracts in one series | Some stocks have two contracts of the same series on the same day (the Sep-2025 Thursday→Tuesday switch; relabelled masters). The contract that traded on the rebalance day is used, a sibling only on days it has no price. Before this rule (fixed 2026-09-28) the choice depended on database row order: 12 holdings in 5 cycles differed between PC and VPS, and the 1x NAV by 0.3% |
| Expiries | 299 series, one per month; holiday shifts (e.g. 2008-12-25 → 24) and the relabelled June-2023 series handled |
Expiry dates come from the stock futures themselves: a date on which at least 10 stock-futures contracts last traded, with holiday-shifted last days. Series less than 3 days apart are merged (June 2023: "29-Jun" until 06-27, "28-Jun" on expiry day). Taking the dates from index-future names was wrong: that also matched NIFTYNXT50 and sector futures, and the stray duplicate-expiry masters.
Earlier engine checks (2010–2013 test window)
| Check | Result |
|---|---|
Leverage identity on the first day of each cycle: ret_L = L·ret_1 − (L−1)·r·days/365 | exact (error 0 on 47 cycles). Later in a cycle the effective leverage drifts until the next reset, by design |
| 1x futures vs an equal-weight spot total-return series with the same members and expiries (built for the check only) | daily correlation 0.996. 1x futures + collateral trailed spot TR by ~3.0% a year: ~1.2% roll costs, the rest the stock-futures premium over the NIFTY-implied rate |
| Split / bonus inside a cycle | 53 scaled, 1 not confirmed by the future (left unscaled) |
F&O effect study (fo_effect.py)
Question (2026-09-27, from the user's observation): does a stock lose momentum once it enters F&O?
backend/nseeod/fo_effect.py builds equal-weight cash indices, total return, base 1000, rebalanced at every
monthly expiry (the same calendar as the leverage study). Members are taken as they were at each expiry.
| Code | Members at each expiry |
|---|---|
FO-EW-EQ | Stocks with a stock future in the next series (the leverage study's list) |
NONFO-N500-EW | NIFTY 500 members (idx_members_on) minus F&O stocks, from 2003 (tracker within ±15 from then) |
NONFO-T750-EW | Own top 750 mainboard stocks by 6-month median traded value (≥ 120 of 126 days traded; ETFs excluded) minus F&O, from Nov 2001 |
FO-NEW12-EW | F&O stocks whose futures started within the last 12 expiries |
SPREAD-FO-N500 / SPREAD-FO-T750 | Daily long FO-EW-EQ, short the non-F&O index |
Why no official "NIFTY 500 + Smallcap 250 = 750": Smallcap 250 is part of NIFTY 500 (ranks 251–500), so the union is ~500 stocks. The official 750-stock index is NIFTY Total Market (NIFTY 500 + Microcap 250), which the constituent tracker does not have. T750 stands in for it.
Method:
- Buy-and-hold inside a cycle (weights drift), back to equal weight at the next expiry.
- A stock with no row on a day returns 0; delisted stocks keep their last close until the next rebalance.
- Prices come from
eq_store.get_returns_panel(adjust='total')(splits, bonuses, dividends; renames joined). - Two kinds of days are skipped (return 0):
- 1,304 stock-days with an unresolved
REVIEWcorporate action: mostly splits and bonuses before 2006 that NSE's records lack (ITC 2005-09-21 showed −93%), and demergers without a factor. - Returns across a gap of more than 31 days: e.g. GECALSTHOM's last row in 2002 joined to AREVA T&D in 2008 gave +4,956%.
- 1,304 stock-days with an unresolved
Event study (ci_event):
- Events: every entry into the F&O list after the Nov 2001 launch batch (576), and every exit (399). A gap of up to 2 expiries out of the list is not counted as an exit.
- Event date: the first (last) day any of the stock's futures traded. Cycle 0 is the expiry cycle containing that date.
- Abnormal return (AR): per cycle from −24 to +24, the stock's total return minus the same-cycle return of
NONFO-T750-EW(or ofNONFO-N500-EW). - Reported: the average AR summed over windows (CAR) with t-stats, overall and per era.
Results (local, 2026-09-27)
| Series | From | Final NAV | CAGR | Vol | Sharpe | Max DD |
|---|---|---|---|---|---|---|
F&O stocks FO-EW-EQ | Nov 2001 | 34,460 | 15.3% | 23.7% | 0.50 | −71.8% |
| Non-F&O NIFTY 500 | Jan 2003 | 72,198 | 19.8% | 20.0% | 0.75 | −69.7% |
| Non-F&O top 750 | Nov 2001 | 90,040 | 19.9% | 23.0% | 0.68 | −77.1% |
| F&O new entrants (≤ 12 expiries) | Feb 2003 | 14,605 | 12.0% | 28.3% | 0.48 | −77.2% |
| Spread F&O − non-F&O NIFTY 500 | Jan 2003 | 482 | −3.0% | 10.9% | — | −59.0% |
| Spread F&O − non-F&O top 750 | Nov 2001 | 337 | −4.3% | 11.7% | — | −70.8% |
Entries (576), abnormal return vs non-F&O top 750:
| Window (expiry cycles) | −24..−1 | −12..−1 | 0 | +1..+12 | +13..+24 |
|---|---|---|---|---|---|
| Mean CAR | +41.1% (t 13.7) | +17.5% (t 8.5) | −0.4% | −3.8% (t −2.4) | +0.9% (t 0.6) |
The pattern holds against the NIFTY 500 benchmark (+46.7% before, −3.5% in +1..+12) and in every era. The post-entry lag was largest in 2014–19 (−12.5% vs top 750 in +1..+12).
Exits (399): −19.9% in the 24 cycles before (t −6.6), and a further −5.8% in +1..+12 (t −2.4). Stocks removed from F&O keep lagging.
Reading:
- The momentum happens before entry, not after. NSE admits stocks on market cap and turnover, so they join after a large run-up (+41% over the benchmark in 2 years).
- After entry the excess return stops, and turns slightly negative in the first year (−3.8%, significant). This is consistent with the observation, but it is "momentum ends" more than "momentum reverses".
- As a group, the F&O stocks trailed the non-F&O stocks by ~4.5 points a year over 25 years, at similar volatility. This is partly a size effect: non-F&O stocks are smaller.
- Limits:
- equal weight;
- no costs;
- delisted stocks keep their last price;
- the unreviewed pre-2006 corporate actions are skipped, not applied;
- the NIFTY 500 tracker is approximate (±15 members) before 2020.
python -m nseeod.fo_effect build [--apply] [--cost-bps 0]
python -m nseeod.fo_effect events [--type entry|exit]
Usage
Run from backend/:
python -m nseeod.custom_index build [--leverage 1,2,3] [--cash-rate implied|6.5] [--cost-bps 5] [--from --to] [--apply]
python -m nseeod.custom_index stats
build fully recomputes the ci_* rows for those series (derived data). The Bhav Downloader recomputes them
after each run, once the tables exist (BhavProcessor._update_custom_index).
| Table | Content |
|---|---|
ci_master | Series code (FO-EW-FUT-1X …), parameters (params JSON), stats JSON (metrics + calendar years) |
ci_level | Daily nav, daily_ret, n_members, gross_exposure, cash_rate |
ci_holding | Per expiry and series: stock, contract, entry price, quantity, weight (cash series: stock and weight only) |
ci_event | F&O entry / exit events: per event and cycle −24..+24, stock return and both benchmarks' returns |
API (app/api/custom_index.py): GET /api/custom-index/series?study=leverage|fo_effect, /levels?codes=, /holdings?code=&date=, /events?type=entry|exit&bench=a|b, /events/list.