Building a Mechanical Trading System

Two traders watch NIFTY move above an earlier high. One calls it convincing. The other waits because it looks stretched. A mechanical trading system gives a third reader clear conditions for the decision: when to enter, how much exposure to accept, when to ask for an exit, and what to do when information is missing.

Systematic means decisions follow written rules. Automated means software carries out part or all of the work. You can check a completed candle against a written threshold and place the order by hand. A program can do the same check and send the order. Coding changes who does the steps. It does not make the idea profitable, and it does not make the order fill.

Discretionary trading leaves some choices to judgment in the moment. You can still review it, if you wrote down what you saw, what you decided and why, at the time. The hard part is rebuilding a judgment nobody recorded, once you already know what happened next. A hybrid is fine to study too, if it says which decisions follow rules and where a person may step in.

Another reader should reach the same signal from the same information, and be able to explain any difference in the result. Judgment does not disappear. It moves into choosing the rules, the data, the fill assumptions and the review tests. Written rules make a deviation visible. They do not guarantee discipline or remove emotion. The previous lesson, on emotional control, is where those deviations start.

Specimen 0.1 is an untested classroom idea with invented prices, times and budgets. Its signal is a completed five-minute NIFTY 50 index candle that closes strictly above the previous session’s high: here 25,020 against 25,000. Equal does not count. The index gives a bullish signal. The instrument to act on it is a bought Call (CE), and its premium is a separate price.

The invented contract is an NSE NIFTY index Call with strike 25,000 and expiry 13 October 2026. Its listing is assumed, not checked against a real exchange file. A real study must confirm that exact expiry and strike in the dated contract file, with the quantity for that contract. Never quietly swap in another expiry. The example uses 65 units per lot only as a teaching quantity. It does not state the current or past exchange lot size.

The specimen looks only at the first qualifying completed bar between 10:00 IST and a 15:00 IST cutoff. If data, quote filters or funding fail at that moment, record no entry for the day and do not hunt for a later signal. If an order is sent, allow no second attempt, filled or not. Before sending, you need a known contract, a complete index bar and a fresh two-sided option quote. The invented quote is bid ₹99 and ask ₹100, with 65 units shown at the ask. The classroom filters allow a spread of at most ₹1 and a quote no older than two seconds. One chance a day is an arbitrary paper-study choice, not a rule for all systems.

The entry is a buy limit at ₹100 per unit. If it is still unfilled after 30 seconds, request cancellation, then reconcile confirmations and any fills that happened meanwhile. After a confirmed entry, an option bid at or below ₹60 triggers an exit request. So does the 15:00 cutoff, whichever comes first. There is no profit target and no planned overnight hold. The exit is a sell limit at the available bid for the confirmed open quantity. An unfilled or rejected exit goes to incident handling and a supervised position check. A requested time or price cannot promise the position is closed.

Size the position on the premium you commit, not on the planned stop loss. At an assumed ₹100 fill, 65 units cost ₹6,500. Add an invented ₹100 round-trip cost reserve and you get ₹6,600. An allowance of ₹10,000 permits floor(10000 ÷ 6600) = 1 lot. Assumed available cash of ₹7,000 also permits 1. Round down and take the smaller answer. If either limit cannot cover one lot, the answer is zero lots. These budgets are teaching inputs, not suggested account limits. The Start lesson on risk management shows how to set your own.

A fill at the planned ₹60 exit gives a net loss of ₹2,700: ₹40 per unit across 65 units, plus ₹100 of costs. That is a planned outcome, not a ceiling on loss. A gap, a missing bid or an unfilled order can make the exit worse, and the whole premium can be lost. The ₹6,600 includes an estimated reserve, so real costs can differ. A futures or short-option system needs its own margin, obligation and portfolio-risk treatment. This premium-only sum does not cover it.

A backtest is a simulation of past decisions and results under a stated rule version, dataset and fill model. It asks what might have happened under those assumptions. It is not a record of what a trader earned. Point-in-time data means using only what was actually available at each simulated decision, including the contract definitions of that day, and not figures corrected or published later.

In our invented sequence, the five-minute index bar completes at 10:00:00 IST. Its final value arrives at 10:00:02, and the signal is checked at 10:00:03. That check is not yet an entry decision. The option quote arrives at 10:00:04 with ask ₹100 and 65 units shown. Only then can the other entry checks run and an order go out. The model assumes a fill at 10:00:05, and only if the price and the quantity are still there when the order reaches the market. These timings are invented, not measured data-feed delays. A signal known at 10:00:02 cannot justify a buy at 09:59:59.

Keep three things apart: the index input, the option quote, and the execution report or assumed fill. A last traded price can be old. It is neither a current ask for buying nor a bid for selling. Displayed depth shows quantity at one moment. It promises no queue place and no later fill. A midpoint fill needs a reason. NSE separates snapshots from levels of order-book depth, and a dataset cannot supply detail it never recorded.

An open-high-low-close bar records four prices, not the order they came in. If an option bar reaches both the stop level and the target, the bar cannot tell you which came first. And a price touched is not a price filled. Use finer data, a conservative rule fixed in advance, or report both outcomes. Otherwise mark the case unresolved. Picking the favourable path after seeing the result adds hindsight to the test.

Keep expired contracts, and the set of contracts that were eligible on each test date. Today’s option chain cannot rebuild yesterday’s choices. Use the session calendar, expiry rules and quantity in force then. For stock options you also need the corporate-action adjustments. Treat stale or missing inputs as unavailable, and record what you left out. An exit with missing data stays unresolved exposure in the model. It does not become a convenient zero or vanish from the sample.

Start with earlier history to build the idea and find errors. Use later history to choose between recorded versions and to see whether costs, fill assumptions or small threshold changes alter the answer. All of that is research and selection. Before you open a final reserved block, freeze the full specification and the test plan. Then report its losses, rejected signals, missing data and failures along with the good fills.

Out-of-sample means the data had no influence on the choices that selected the tested version. A later date alone does not give you that. If the reserved result talks you into changing the entry, exit, filter or size, that block becomes research input for the new version. Running the new rule on the same block is another research comparison. It is no longer fresh, unseen evidence.

These eight closed Call round trips are a separate invented accounting sample. They are not market data and not results from Specimen 0.1. Every entry is assumed at ₹100 for 65 units. Exits are ₹120, ₹85, ₹115, ₹80, ₹110, ₹90, ₹125 and ₹78, and each round trip carries an invented ₹100 total charge. The ledger teaches arithmetic, not how the candidate performs.

For each row, the gross result is (exit − entry) × quantity, and the net result takes off costs once. The eight gross results add up to a ₹195 gain. Total assumed costs are ₹800, which leaves a ₹605 net loss. In a real test, define brokerage, exchange and statutory charges for the dates and trades involved. Also model spread and slippage, the gap between your reference price and the price you got. If your bid and ask prices already include crossing the spread, do not charge that spread a second time.

A signal, an accepted order and an executed trade are three different events. An order can be acknowledged, rejected, left pending or filled in parts. A stop condition turning true creates an instruction in your system. It does not reserve an exit price. NSE describes partial order matching, and Kite Connect’s documentation separates placing an order from executing it and shows filled and pending quantities. The exact labels on your terminal depend on your provider.

Here is a separate invented execution record with three already-bought Call lots: 195 units at 65 units a lot. An exit request for all three lots fills one lot, leaving 130 units open. You request cancellation of the two outstanding lots. Before it is confirmed, another 65 units fill. The confirmed cancellation then removes the last 65-unit exit order. Total exited quantity is 130, and one lot of 65 is still open. Cancelling an order has not sold the position. This case uses whole-lot fills and is separate from the one-lot candidate and the eight accounting rows.

Before you send another instruction, reconcile the order log, the trade confirmations and the current position. A connection timeout can leave the order state unknown, and retrying as if the first request failed can double your exposure. Pause new entries while you resolve the mismatch, keep the timestamps, and decide who checks the position and how. Switching off new orders does not manage an existing position on its own. An emergency may need action under your written incident procedure.

Check available funds and open obligations again at the moment of decision. An earlier allowance does not prove that premium cash or margin is still free after other positions, charges or changed requirements. A roll closes or reduces one contract and opens another, each with its own execution, funding and costs. Following rules by hand and routing orders by program also have different operating needs. Before you automate anything, check the current SEBI, exchange and broker rules for that specific route. A classroom checklist is not permission to run a bot.

Review the process and the outcomes separately. Was the input available at the recorded time? Did the chosen contract meet the specification? Were quantities, fills and costs reconciled? Did any manual decision depart from the rule? A profitable exception is still an exception, and a correctly executed loss is still a loss. Keeping those apart lets a reviewer diagnose the process without rewriting the result.

Keep every rule version, its data, its tested assumptions and the alternatives you looked at. A wider stop or a different entry cutoff starts a new research version, with a reason and an effective date. Keep the earlier history, including the final block that prompted the change. An urgent safety step during an incident may be needed. Record the circumstances and the orders. Do not insist that nothing may ever change mid-trade, and do not quietly relabel the exception as the tested rule.

Choose your review and pause conditions in advance, and look into serious failures as they happen. The conclusion may be rejection, data repair, more paper observation or a new research version. Explain the decisions and the exposure, including why the system did nothing. A good research result alone does not settle a separate decision about capital or going live. The next lesson, dual-journaling, shows how to link what you intended to what you did.