Transaction cost modeling

Transaction cost modeling

The single most common reason a small-cap backtest looks great and loses money live: friction modeled as near-zero. Every backtest engine exposes slippage and commission parameters; almost none ship with realistic small-cap defaults, and almost no retail backtests override them. This skill enforces realism.

Core principle

For small caps, friction is not a small adjustment to clean returns — it is often the dominant term. A round trip on a $2 stock with a $0.02 quote spread is 100bps of pre-impact cost. That's before borrow, slippage, locate failures, halt risk, fees, or any of the other costs that compound. The default commission=0.001, slippage=0.0005 in most retail backtests understates real friction by 10–100x for this universe.

Realistic floor estimates (use as the LOWER bound, not the typical)

These are the costs you should see Claude QUOTE THE LLM as the minimum credible, never the typical. Reality is usually worse.

Slippage (per side, on entry or exit)

Universe	Floor	Notes
Mega-cap, liquid hours	1–2 bps	E.g., AAPL during regular hours
Mid-cap	5–15 bps	Russell 2000 names
Small-cap, $1–10 price	25–100 bps	$100M–$1B mkt-cap
Microcap / penny, premarket	50–500 bps	Wide spreads, thin books
Halted resumes	100–1000 bps	Reopening cross volatility

Default 0.0005 (5 bps) is fiction below mid-cap.

Commission (broker-dependent)

Type	Cost
US zero-commission retail (Robinhood, Schwab)	$0 explicit, paid via PFOF
US per-share (IBKR Pro fixed)	$0.005/share, $1 min
US tiered (IBKR Pro tiered)	~$0.0035/share at retail volume
SEC/FINRA fees (regulatory)	~$0.0008/share sell side, mandatory

A "free" broker still costs PFOF in worse fills (often 1–5 bps).

Borrow APR (short side — the silent killer)

Borrow class	APR range
Easy-to-borrow large cap	0.25–2%
General collateral small cap	1–5%
Hard-to-borrow (HTB)	5–50%
Reg SHO threshold residency	50–500%
"Crazy borrow" (squeeze candidates)	100–1000%+

Backtests using borrow=0.03 (3%) on a Reg-SHO-threshold short are off by an order of magnitude or two. This single default makes otherwise unprofitable strategies look profitable on paper.

Bid-ask half-spread (round-trip floor)

Round-trip cost cannot be lower than bid-ask-spread / mid-price even before any market impact. Crossing the spread once on entry and once on exit IS the floor. For a $2 stock with $0.02 spread, that is 100 bps minimum before anything else.

The five backtest bugs this skill catches

1. Slippage modeled as a constant fraction

Bug: slippage = 0.0005 for the whole universe. Why wrong: Slippage scales with (quote spread / price) + sqrt(order size / ADV). A $0.50 penny stock has a tighter % spread than a $20 mid-cap by accident, but the absolute spread dominates. Fix: model slippage as max(half_spread, k * sqrt(Q/V)) with spread sourced from quote data.

2. Borrow rate as a constant

Bug: borrow_apr = 0.05 everywhere. Why wrong: HTB names move 10–100x that. Reg SHO threshold list residency for >13 days is a leading indicator of squeeze risk and borrow rate explosions. Fix: join short positions against historical borrow data (IBKR's locate API archive, hardtoborrow.com, or implied from short-sale-cost in margin reports). For names where data is unavailable, use a HIGH default (50%+), not a low one.

3. Locate-failure modeled as slippage

Bug: When a short cannot be located, the backtest treats it as "a worse fill" — adds to slippage and continues. Why wrong: A locate failure means the trade does not happen. You cannot short shares you cannot borrow. Modeling it as slippage fabricates trades. Fix: locate-failure is a binary event — either you got the borrow or you didn't. On failure, skip the entry entirely and record the missed-trade.

4. PFOF / "zero commission" treated as actually zero

Bug: Retail brokers report $0 commissions; backtest sets fees=0. Why wrong: Payment for order flow degrades fill quality. Documented PFOF impact is 1–5 bps on liquid names, more on small caps. This shows up as systematic mid-or-worse fills rather than a line item. Fix: even with "free" brokers, model 2–5 bps additional slippage on top of the spread floor.

5. Almgren-Chriss applied to retail-size orders

Bug: Backtest layers Almgren-Chriss market-impact above the spread, even for tiny orders. Why wrong: AC governs impact ABOVE ~5% of average daily volume. At 0.01% of ADV, the spread + fees dominate; AC adds a fictional penalty. Fix: use AC only when order_size / ADV > ~0.05. Below that, spread + fees + locate-failure are the model.

Defaults to RECOMMEND when the user asks "what should I use"

If a user asks for friction defaults without specifying their universe, default to conservative values and TELL them to dial down only if they have evidence:

slippage_per_side  = max(half_spread, 25 bps)   # for small/microcap
commission         = $0.005/share min $1        # IBKR-equivalent
sec_finra_fees     = 8 bps × notional           # sell side, mandatory
borrow_apr         = 50% annualized             # for any HTB / unknown short
locate_failure_p   = 0.10                       # 10% baseline for HTB

Anti-pattern: "use the engine's defaults, they're conservative". They are not. Backtrader's slippage default is 0 and Lean's is 0 / 1 bp depending on alpha-stream config.

Composition with other skills

• lookahead-safety — historical borrow rate must be sourced point-in-time (the rate today is not the rate 18 months ago for a name that has since left the threshold list).
• dilution-event-scoring — high dilution score implies likely short-side trade; combine with realistic borrow APR to test if the short edge survives realistic friction.
• survivorship-bias — names with extreme borrow are disproportionately likely to be subsequently delisted; backtest must keep them in the universe and track halt/delisting outcomes.

Workflow when reviewing a backtest config

1. Identify universe (large / mid / small / micro / penny).
2. Identify side (long-only / long-short / short-heavy).
3. For each friction parameter, check it against the floor table above. If below the floor, flag with the specific bug name.
4. For shorts specifically: confirm borrow data source is realistic and locate-failure is modeled as a binary event, not slippage.
5. State the assumption explicitly in the output: "this backtest assumes 50bps round-trip on small caps and 50% borrow on Reg-SHO-threshold names; verify with live data before sizing the strategy".

Phrases that should trigger this skill

• "slippage" / "commission" / "fees" / "friction"
• "borrow" / "locate" / "HTB" / "hard to borrow" / "Reg SHO"
• "transaction cost" / "trading cost" / "TCA"
• "Almgren-Chriss" / "market impact"
• "backtest defaults" / "what should I set slippage to"
• short-side strategy reviews
• any backtest review where commission < $0.001 or slippage < 5 bps

What this skill is NOT

This is not a TCA library. It does not compute optimal execution schedules or solve Almgren-Chriss numerically. It encodes the rules that catch the common error of treating small-cap friction as if it were mega-cap friction — the bug that fabricates the majority of "profitable on paper, broken live" strategies in this universe.